CN117520905B - Anti-fact fault data generation method based on causal intervention - Google Patents
Anti-fact fault data generation method based on causal intervention Download PDFInfo
- Publication number
- CN117520905B CN117520905B CN202410005964.9A CN202410005964A CN117520905B CN 117520905 B CN117520905 B CN 117520905B CN 202410005964 A CN202410005964 A CN 202410005964A CN 117520905 B CN117520905 B CN 117520905B
- Authority
- CN
- China
- Prior art keywords
- patch
- real
- fault data
- generator
- fault
- 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.)
- Active
Links
- 230000001364 causal effect Effects 0.000 title claims abstract description 92
- 238000000034 method Methods 0.000 title claims abstract description 35
- 238000012549 training Methods 0.000 claims abstract description 6
- 239000013598 vector Substances 0.000 claims description 42
- 230000008569 process Effects 0.000 claims description 15
- 238000012545 processing Methods 0.000 claims description 15
- 238000009826 distribution Methods 0.000 claims description 12
- 238000013507 mapping Methods 0.000 claims description 11
- 238000005457 optimization Methods 0.000 claims description 8
- 230000008485 antagonism Effects 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 claims description 2
- 238000005259 measurement Methods 0.000 claims description 2
- 238000003745 diagnosis Methods 0.000 abstract description 9
- 230000007246 mechanism Effects 0.000 abstract description 4
- 230000006870 function Effects 0.000 description 10
- 230000007613 environmental effect Effects 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 239000012634 fragment Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011897 real-time detection Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0475—Generative networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/094—Adversarial learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
The invention relates to the technical field of anti-fact fault data generation, in particular to an anti-fact fault data generation method based on causal intervention. Firstly, describing a generation mechanism of fault data through a structural causal model, considering that the fault data consists of causal features and non-causal features, and only the causal features can influence the judgment of fault types. And then decoupling through a CycleGAN network to generate causal features and non-causal features. And introducing causal relation loss and characteristic information contrast loss in the CycleGAN network to constrain the model, and further reserving causal factors and intervening non-causal factors. Training a generator and a discriminator in the CycleGAN network to obtain an optimal generator, and generating anti-fact fault data through the optimal generator. The invention can generate high-quality anti-reality fault data through the proposed network model, and improve the precision of fault diagnosis.
Description
Technical Field
The invention relates to the technical field of anti-fact fault data generation, in particular to an anti-fact fault data generation method based on causal intervention.
Background
At present, a new path is opened up for intelligent fault diagnosis based on a deep learning method, and the real-time detection of the mechanical equipment faults is realized by extracting the characteristics in fault data. However, in the deep learning diagnostic model, the collected fault data are also quite different due to different working environments of the mechanical equipment, so that the robustness and accuracy of the same diagnostic model using the fault data are also problematic, and the following main reasons are:
1. due to the difference of working conditions of mechanical equipment and data acquisition methods, fault data acquired for fault diagnosis contain a large number of environmental features which have no causal relation with fault types but affect the diagnosis accuracy of a diagnosis model; and the fault data collected under different environments usually have different distribution characteristics, so that the generalization capability of the same diagnosis model on the fault data is reduced, and finally, the prediction accuracy is reduced.
2. The number of fault data which can be obtained at present and does not contain excessive noise is small, so that the diagnosis model trained by using limited data quantity is difficult to fully mine key features for judging fault types, and further the diagnosis precision of the diagnosis model is reduced.
It follows that further research is now required in terms of acquisition of fault data.
Disclosure of Invention
In order to avoid and overcome the technical problems in the prior art, the invention provides a causal intervention-based anti-reality fault data generation method. The invention can decouple the causal feature and the non-causal feature, and generate diversified anti-reality fault data through the intervention of the non-causal feature, thereby reducing or eliminating the influence of the non-causal feature on the fault type judgment.
In order to achieve the above purpose, the present invention provides the following technical solutions:
a method for generating anti-fact fault data based on causal intervention comprises the following steps:
s1, acquiring original working condition fault data and target working condition fault data; extracting real fault characteristics in original working condition fault data, inputting the real fault characteristics into a generator in a generating countermeasure network, performing characteristic decoupling on the real fault characteristics, and decoupling the real fault characteristics into real causal characteristics with causal relation between fault types and real non-causal characteristics without causal relation between the fault types; the real causal features generate corresponding anti-facts causal features through a generator, and the anti-facts causal features form corresponding anti-facts fault features;
s2, inputting the counter-facts fault characteristics and the real fault characteristics in the target working condition fault data into a discriminator in a generating countermeasure network at the same time, performing error optimization training on the discriminator and the generator to obtain an optimal generator, and forming a mapping for generating the real fault characteristics in the target working condition fault data from the real fault characteristics in the original working condition fault data in the optimal generator;
s3, inputting a real fault sample in the original working condition fault data into an optimal generator, and generating corresponding anti-facts fault data through mapping.
As a further scheme of the invention: true faults in original working condition fault dataAfter the characteristics are input into a generator in a generating countermeasure network, the generator firstly performs characteristic decoupling on the real fault characteristics, namely, real vibration signals of the real fault characteristics are divided into a plurality of real vibration signals in time sequenceKEach real patch corresponds to a position in the real vibration signal, and each real patch is respectivelyz 1 、z 2 、…、z k 、…、z K The method comprises the steps of carrying out a first treatment on the surface of the Wherein,z 1 representing the 1 st actual patch to be applied,z 2 representing the 2 nd real patch;z k represent the firstkA real patch;z K represent the firstKA real patch;
real causal features with causal relation between each real patch representation and the fault type;
inputting the real patches into a generator, sequentially passing through an input layer, a network of each layer and an output layer in the generator, and finally outputting corresponding anti-facts patches in the generator, wherein each anti-facts patch is respectivelyThe method comprises the steps of carrying out a first treatment on the surface of the Wherein (1)>Representing the 1 st counterfactual patch output by the generator, i.ez 1 A counterfactual patch generated in the generator; />Representing the 2 nd counterfactual patch output by the generator, i.e.z 2 A counterfactual patch generated in the generator; />Represents the first output of the generatorkCounter-facts patches, i.ez k A counterfactual patch generated in the generator; />Represents the first output of the generatorKInverse of each otherFact patches, i.e.z K A counterfactual patch generated in the generator; the anti-facts patches represent corresponding anti-facts causal features, and each anti-facts patch is sequentially arranged and combined according to the positions of the corresponding real patches in the real vibration signals, so that anti-facts fault data generated by the generator through the original working condition fault data are formed.
As still further aspects of the invention: the generation of the antagonism network adopts a CycleGAN network, and the antagonism loss function of the CycleGAN network is expressed as follows:
；
wherein,representing a fight loss function for the CycleGAN network;Ga representation generator;Da representation discriminator;Xrepresenting a sample set of raw operating condition fault data,xrepresentation ofXA true fault sample in (a); />Data distribution representing original operating mode fault data;E x representation ofXExpectations of the true fault samples in (a);D(. Cndot.) represents the discriminant function of the CycleGAN network;G(. Cndot.) represents the generation function of the CycleGAN network;Ya sample set of target operating condition fault data is represented,yrepresentation ofYA true fault sample in (a); />Data distribution representing target operating condition fault data;E y representation ofYIs a true failure sample.
As still further aspects of the invention: in a real vibration signal of a real fault feature with a time stamp in original working condition fault data, the relation between real patches at different moments is inconsistent, and causal relation between different real patches is calculated by dot product operation;
in the original working condition fault data, for a given real patch expressed by a vectorz k It is associated with a real patchz i The correlation relationship between them is expressed as:
；
wherein P is k (i) Representing a real patchz k And a real patchz i The correlation between them evaluates the score, i.e. characterizes the real patchz k And a real patchz i A causal relationship distribution between the two;exp(. Cndot.) expressed in terms of natural constantseAn exponential function of the base;urepresenting the super-parameters;z j the first generated in the true vibration signal representing the original condition fault datajA real patch, j=1, …,K，Krepresenting the total number of real patches;Trepresenting a matrix transpose operation;
solving real patches in original working condition fault dataz k And all other real patches, and is noted asP k ；
Similarly, in the counterfactual fault data, for a given counterfactual patch represented by a vectorIt is associated with a counterfactual patch>The correlation relationship between them is expressed as:
；
wherein Q is k (i) Representing a counterfactual patchAnd counterfactual patch->The correlation between them evaluates the score, i.e. characterizes the counterfactual patch +.>And counterfactual patch->A causal relationship distribution between the two; />Representing the first of the anti-facts vibration signals formed in the anti-facts fault datajA counter fact patch;
solving the counterfactual patch in the counterfactual fault dataCorrelation with all other counterfactual patches and is noted asQ k ；
Via JS divergence measureP k AndQ k similarity between the two, the measurement formula is as follows:
；
wherein,representation ofP k AndQ k similarity between; />Representation ofQ k For a pair ofP k KL divergence of (2); />Representation ofP k For a pair ofQ k KL divergence of (2);
by minimizingI.e.The real patch can be further constrainedz k And corresponding anti-facts patchIntegrity of causal features in between;
based on minimizationThe similarity between all the real patches and the corresponding anti-facts patches is solved through JS divergence, and the solved result can be used as a causal relation loss, wherein the causal relation loss is expressed as follows:
；
wherein,representation generatorGIs lost in causal relationship in (a).
As still further aspects of the invention: acquiring characteristic information extracted from different layers of networks in a discriminator, wherein the discriminator is sharedLA layer network in whichlLayer common outputN l With dimensions of 1XM l Is used for the feature vector of (a),M l representing the length of the feature vector,lthe value ranges from 1 toLThe method comprises the steps of carrying out a first treatment on the surface of the The number of the real fault characteristics and the counter-fact fault characteristics generated in the same layer network of the discriminator is the same;
the loss between the counter fact feature vector generated by the counter fact fault data in the layers of the network of the discriminator and the real feature vector generated by the target working condition fault data in the layers of the network of the discriminator is called layer contrast loss, and the layer contrast loss is expressed as follows:
；
wherein,L con l h,, representing the counterfactual fault data at the arbiterlAdverse events generated in a hierarchical networkThe real characteristic vector and the target working condition fault data are in the first of the discriminatorslThe third layer is a layer network of the true feature vectorshLoss of individual layer contrast;representing the counterfactual fault data at the arbiterlGenerated in a layer networkhA counter-fact feature vector is used to determine,hthe value ranges from 1 toN l ，N l Representation of the discriminantlThe total number of the counter fact feature vectors or the real feature vectors generated in the layer network;representing the counterfactual fault data at the arbiterlGenerated in a layer network except->A feature vector set composed of all other inverse feature vectors except for the others; />Representation->The number of the middle-inverse fact feature vectors; />Representation ofThe first of (3)rA counter-fact feature vector is used to determine,rthe value range is 1 to->；/>Representation->The first of (3)fA plurality of inverse fact feature vectors; />Indicating the target working condition fault data in the first discriminatorlA feature vector set formed by all real feature vectors generated in the layer network; />Representation->The first of (3)rTrue feature vectors;ωrepresenting the super-parameters;
the contrast losses of all layers in the discriminator are added to form the total contrast loss of the discriminator, and the total contrast loss is expressed as follows:
；
wherein,representing the total contrast loss value of the discriminator;Lindicating the total number of layers of the network in the arbiter.
As still further aspects of the invention: selecting the anti-facts fault characteristics output by different layers of networks from the generator to perform patch comparison, and obtaining multi-layer patch comparison loss, wherein the comparison process is specifically as follows:
S2A1, in the process of processing original working condition fault data by the generator, obtaining the counterfactual patches output by each layer of network from the generator, and inputting the counterfactual patches into the multi-layer perceptron network one by one, wherein the multi-layer perceptron network outputs the counterfactual patch characteristics corresponding to each layer of network, and the counterfactual patch characteristics are expressed as follows:
；
wherein,representation generatorGFirst, thebLayer network ofb i The true patches are inverse fact patches generated by the generating function;MLPrepresentation ofA multi-layer perceptron; />Representation->After the multi-layer perceptron is input, the corresponding output counterfactual patch features;
S2A2, calculating losses among the anti-fact patch features in the generator layer networks, the positive sample patch features in the anti-fact patch features and the negative sample patch features of the anti-fact patch features to obtain anti-fact patch losses of the anti-fact fault features, wherein the anti-fact patch losses are expressed as follows:
；
wherein,the counterfactual patch loss of original working condition fault data is represented; />Representation->Corresponding negative example patch feature, +.>Representation->Corresponding positive sample patch features;Orepresenting a loss function;Brepresenting the total number of layers of the network in the generator;b I the first of the representation generatorsbThe total number of counterfactual patch features in the layer network;
S2A3, inputting target working condition fault data into a generation countermeasure network, and processing the target working condition fault data according to the processing process of the original working condition fault data in the generator so as to obtain a corresponding real patch; processing the real patch according to the same processing mode as the step S2A1 and the step S2A2 to obtain the counterfactual patch loss corresponding to the target working condition fault data;
S2A4, combining the counterfactual patch loss of the original working condition fault data and the counterfactual patch loss of the target working condition fault data to obtain a multi-layer patch loss of the generator, wherein the multi-layer patch loss is expressed as follows:
；
wherein,a multi-layer patch loss representing a generator; />Representation ofWeight parameters of (2); />The counterfactual patch loss of the target working condition fault data is represented; />Representation->Weight parameters of (c).
As still further aspects of the invention: error optimization refers to simultaneously solving minimum value of fight loss function in CycleGAN network, generatorGMinimum and generator of causal relation losses of (a)GMulti-layer patch loss minima and arbiterDBy repeating the error optimization process and reaching the set stop condition, an optimal generator can be generated, and the mapping of the real fault characteristics in the original working condition fault data to the real fault characteristics in the target working condition fault data can be obtained.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention extracts real causal features related to fault types and irrelevant real non-causal features from real fault samples, and intervenes the non-causal features through decoupling action of a generator to generate anti-fact fault data; and through the error optimization training generator, the accuracy of the anti-fact fault data generated by using the mapping is improved, and the prediction accuracy of the fault type classification model after subsequent training is improved.
2. The fault data of the invention are decoupled into causal and non-causal features, and the anti-fact fault data are generated through independent intervention of the latter. On one hand, causal relation loss is designed, and causal characteristics of original working condition fault data are captured. On the other hand, the contrast loss of the characteristic information is added, and the intervention effect on the non-causal characteristics is improved. The generated anti-facts fault data is helpful to build a robust downstream fault classifier model, and the influence of confounding factors is reduced.
Drawings
Fig. 1 is a schematic diagram of a counterfactual fault data generation model in the present invention.
FIG. 2 is a schematic diagram of a causal relationship between different patches according to the present invention.
FIG. 3 is a schematic diagram of a comparison between different non-causal features in the inventive arbiter.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Referring to fig. 1 to 3, the bearing is taken as an example of the present invention. In the working process of the bearing, the working environment of the bearing is complex and changeable. On the same or different machines, the different fault vibration signals acquired follow different marginal distributions due to changes in environmental factors such as load conditions, temperature, etc. However, in the failure samples, the damage-related characteristics of the components such as the outer ring, the inner ring, the rollers, etc. related to the failure type are not interfered by these environmental factors. I.e. the non-causal features related to environmental factors and the causal features related to fault type are independent of each other, i.e. the independence of the causal mechanisms. The independence of causal mechanisms suggests that intervention on one mechanism does not make changes to the other. And the types and the number of the causal features affecting the fault type are fixed, while the non-causal features can change along with the change of the external environment. The invention aims to separate the causal features and the non-causal features, further reduce or eliminate the influence of the non-causal features on fault type judgment, and improve the accuracy of data.
The anti-facts fault data obtained by the anti-facts reasoning are data which have never appeared. Given real fault data, if the anti-facts fault data generated by the real fault data exist in the real distribution of the real fault samples, the obtained anti-facts fault data is called anti-facts confidence. This process of generating the anti-fact fault data of the anti-fact confidence from the real fault data is referred to as mapping. The mapping process is actually to generate causal features and non-causal features through group decoupling fault data, reserve the causal features, and intervene the non-causal features to ensure the anti-fact confidence.
To achieve the anti-facts confidence, an anti-facts fault data generation model is then built, as shown in fig. 1, mainly divided into two phases, phase one: training a counterfactual generator; stage two: and generating the counterfactual data and diagnosing faults. The generation model takes a CycleGAN network in a generation countermeasure network as a backbone network, and combines a plurality of loss functions to generate the counterfactual fault data.
By arranging various sensors on the fault bearing of which the bearing fault type is determined, the fault bearing is installed in corresponding detection equipment, corresponding working condition data such as load, rotating speed and the like are set, and vibration signals of the fault bearing are detected. And recording fault type and working condition data of the fault bearing and corresponding vibration signals to form fault data. And acquiring real fault data under different environments through multiple times of time, and respectively naming the real fault data as target working condition fault data and original working condition fault data.
After the original working condition fault data are input into a generator in the CycleGAN network, the generator firstly extracts all real fault characteristics and performs characteristic decoupling, the real fault data are divided into a plurality of fragments which are not overlapped with each other, namely real vibration signals of the real fault characteristics are sequentially divided into K patches according to time sequence, the patches are named as real patches, and each real patch corresponds to one position in the real vibration signals. Each real patch representation has real causal features of causal relation with the fault type.
The real patches are input into the generator, each real patch sequentially passes through an input layer, each layer of network and an output layer in the generator to be processed, and finally, the corresponding anti-facts patch is output in the generator. The anti-facts patches represent the corresponding anti-facts and cause-effect characteristics, and the anti-facts fault data generated by the generator through the original working condition fault data can be formed by sequentially arranging and combining the anti-facts patches according to the positions of the corresponding real patches in the real vibration signals.
The aim of the generator is to generate the anti-fact fault data which is as close to the real fault data distribution in the target working condition fault data as possible; the objective of the discriminator is to distinguish the counterfactual fault data generated by the original condition fault data from the target condition fault data. The countermeasures against losses enable the generator to generate counterfactual fault data having specific target operating condition attributes by facilitating the mutual boosting of the generator and the arbiter. Both the generator and the arbiter in the CycleGAN network are typically constructed using neural networks.
Selecting the anti-facts fault characteristics output by different layers of networks from the generator to perform patch comparison, and obtaining multi-layer patch comparison loss, wherein the comparison process is specifically as follows:
and in the process of normally processing the original working condition fault data by the generator, obtaining the counterfactual patches output by the networks of all layers from the generator, and inputting the counterfactual patches into the multi-layer perceptron networks of the two layers one by one, wherein the multi-layer perceptron networks output patch characteristics corresponding to the networks of all layers.
The losses among the patch features in the network of each layer of the generator, the positive sample patch features in the patch features and the negative sample patch features of the patch features are calculated to obtain the counterfactual patch losses of the counterfactual fault features.
Inputting the target working condition fault data into a generation countermeasure network, and processing the target working condition fault data according to the processing process of the original working condition fault data in the generator so as to obtain a corresponding real patch; and processing the real patch according to the same processing mode as the above content to obtain the counterfactual patch loss corresponding to the target working condition fault data.
The counterfactual patch loss of the original operating condition fault data and the counterfactual patch loss of the target operating condition fault data are combined to obtain the multi-layer patch loss of the generator.
On the basis of the multi-layer patch loss, causal relation loss and characteristic information contrast loss are introduced to restrict the causal intervention effect of the generated model, and high-efficiency class-specific characteristic transfer is realized.
As shown in fig. 2, in the real vibration signal of the real fault feature with the time stamp in the original working condition fault data, the relation between the real patches at different moments is inconsistent, and the causal relation between the different real patches is calculated by using dot product operation. And solving the similarity between all the real patches and the corresponding anti-facts patches through JS divergence, wherein the solved result can be used as a causality loss.
Machines and environments mainly contain low-level, operating-specific attributes, i.e., non-causal features. The performance of the arbiter is good or bad in its ability to discriminate against non-causal features. The comparison concept was introduced into the arbiter in order to further improve the performance of the arbiter. The method is characterized in that the network structure of the discriminator is reconstructed, and the characteristic information in the fault data is further extracted by utilizing a plurality of characteristic vectors output by different layers of the network structure.
Acquiring characteristic information extracted from different layers in own network of discriminators, wherein the discriminators are sharedLA layer, wherein the firstlLayer outputN l With dimensions of 1XM l Is used for the feature vector of (a),ltake a value of 1 toLThe method comprises the steps of carrying out a first treatment on the surface of the The number of real fault signatures and anti-facts fault signatures generated in the same layer network of the arbiter are the same.
As shown in fig. 3, the loss between the anti-facts feature vector generated in the respective layers of the arbiter by the anti-facts fault data and the true feature vector generated in the respective layers of the arbiter by the target operating condition fault data is referred to as a layer contrast loss. The contrast losses of all layers in the discriminator are added to form the total contrast loss of the discriminator.
Intuitively, through total contrast loss, the generator may possess greater discrimination, directing the generator to increase the degree of intervention on non-causal features, thereby increasing sample variance.
Error optimization refers to simultaneously solving the minimum value and generator of the fight loss function of the CycleGAN networkGMinimum and generator of causal relation losses of (a)GMulti-layer patch loss minima and arbiterDAnd (3) repeating the error optimization process, stopping iteration when the error reaches a set precision range or reaches corresponding iteration times, and obtaining the mapping of generating the real fault characteristics in the target working condition fault data from the real fault characteristics in the original working condition fault data.
And inputting the original working condition fault data containing the real fault characteristics and the real fault samples of the real fault types corresponding to the real fault characteristics into an optimal generator, and generating corresponding anti-facts fault data through mapping. The generated anti-reality fault data is more real, and the generated data volume can also be determined according to the actual situation. And the generated anti-reality fault data can be mixed with the real fault data and input into a neural network for fault type prediction, so as to train the neural network and improve the prediction accuracy.
The foregoing is only a preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art, who is within the scope of the present invention, should make equivalent substitutions or modifications according to the technical scheme of the present invention and the inventive concept thereof, and should be covered by the scope of the present invention.
Claims (4)
1. The method for generating the anti-fact fault data based on causal intervention is characterized by comprising the following steps of:
s1, acquiring original working condition fault data and target working condition fault data; extracting real fault characteristics in original working condition fault data, inputting the real fault characteristics into a generator in a generating countermeasure network, performing characteristic decoupling on the real fault characteristics, and decoupling the real fault characteristics into real causal characteristics with causal relation between fault types and real non-causal characteristics without causal relation between the fault types; the real causal features generate corresponding anti-facts causal features through a generator, and the anti-facts causal features form corresponding anti-facts fault features;
s2, inputting the counter-facts fault characteristics and the real fault characteristics in the target working condition fault data into a discriminator in a generating countermeasure network at the same time, performing error optimization training on the discriminator and the generator to obtain an optimal generator, and forming a mapping for generating the real fault characteristics in the target working condition fault data from the real fault characteristics in the original working condition fault data in the optimal generator;
s3, inputting a real fault sample in original working condition fault data into an optimal generator, and generating corresponding anti-facts fault data through mapping;
after the real fault characteristics in the original working condition fault data are input into the generator in the generation countermeasure network, the generator firstly performs characteristic decoupling on the real fault characteristics, namely, real vibration signals of the real fault characteristics are divided into time sequences in turnKEach real patch corresponds to a position in the real vibration signal, and each real patch is respectivelyz 1 、z 2 、…、z k 、…、z K The method comprises the steps of carrying out a first treatment on the surface of the Wherein,z 1 representing the 1 st actual patch to be applied,z 2 representing the 2 nd real patch;z k represent the firstkA real patch;z K represent the firstKA real patch;
real causal features with causal relation between each real patch representation and the fault type;
inputting the real patches into a generator, sequentially passing through an input layer, a network of each layer and an output layer in the generator, and finally outputting corresponding anti-facts patches in the generator, wherein each anti-facts patch is respectivelyThe method comprises the steps of carrying out a first treatment on the surface of the Wherein (1)>Representing the 1 st counterfactual patch output by the generator, i.ez 1 A counterfactual patch generated in the generator; />Representing the 2 nd counterfactual patch output by the generator, i.e.z 2 A counterfactual patch generated in the generator; />Represents the first output of the generatorkCounter-facts patches, i.ez k A counterfactual patch generated in the generator; />Represents the first output of the generatorKCounter-facts patches, i.ez K A counterfactual patch generated in the generator; the anti-facts patches represent corresponding anti-facts causal features, and the anti-facts patches are sequentially arranged and combined according to the positions of the corresponding real patches in the real vibration signals to form anti-facts fault data generated by the original working condition fault data through the generator;
the generation of the antagonism network adopts a CycleGAN network, and the antagonism loss function of the CycleGAN network is expressed as follows:
wherein,representing a fight loss function for the CycleGAN network;Ga representation generator;Da representation discriminator;Xrepresenting a sample set of raw operating condition fault data,xrepresentation ofXA true fault sample in (a); />Data distribution representing original operating mode fault data;E x representation ofXExpectations of the true fault samples in (a);D(. Cndot.) represents the discriminant function of the CycleGAN network;G(. Cndot.) represents the generation function of the CycleGAN network;Ya sample set of target operating condition fault data is represented,yrepresentation ofYA true fault sample in (a); />Data distribution representing target operating condition fault data;E y representation ofYExpectations of the true fault samples in (a);
in a real vibration signal of a real fault feature with a time stamp in original working condition fault data, the relation between real patches at different moments is inconsistent, and causal relation between different real patches is calculated by dot product operation;
in the original working condition fault data, for a given real patch expressed by a vectorz k It is associated with a real patchz i The correlation relationship between them is expressed as:
wherein P is k (i) Representing a real patchz k And a real patchz i The correlation between them evaluates the score, i.e. characterizes the real patchz k And a real patchz i A causal relationship distribution between the two;exp(. Cndot.) expressed in terms of natural constantseAn exponential function of the base;urepresenting the super-parameters;z j the first generated in the true vibration signal representing the original condition fault datajA real patch, j=1, …,K，Krepresenting the total number of real patches;Trepresenting a matrix transpose operation;
solving real patches in original working condition fault dataz k And all other real patches, and is noted asP k ；
Similarly, in the counterfactual fault data, for a given counterfactual patch represented by a vectorIt is associated with a counterfactual patch>The correlation relationship between them is expressed as:
wherein Q is k (i) Representing a counterfactual patch->And counterfactual patch->The correlation between them evaluates the score, i.e. characterizes the counterfactual patch +.>And counterfactual patch->A causal relationship distribution between the two; />Representing the first of the anti-facts vibration signals formed in the anti-facts fault datajA counter fact patch;
solving the counterfactual patch in the counterfactual fault dataCorrelation with all other counterfactual patches and is noted asQ k ；
Via JS divergence measureP k AndQ k similarity between the two, the measurement formula is as follows:
wherein (1)>Representation ofP k AndQ k similarity between; />Representation ofQ k For a pair ofP k KL divergence of (2); />Representation ofP k For a pair ofQ k KL divergence of (2);
by minimizingCan further restrict the real patchz k And corresponding counterfactual patch->Integrity of causal features in between;
based on minimizationThe similarity between all the real patches and the corresponding anti-facts patches is solved through JS divergence, and the similarity can be used as a causal relation loss, and the causal relation loss is expressed as follows:
wherein (1)>Representation generatorGIs lost in causal relationship in (a).
2. The method for generating anti-facts fault data based on causal intervention according to claim 1, wherein the feature information extracted from different layers of networks in the discriminators is obtained, and the discriminators are sharedLA layer network in whichlLayer common outputN l Each dimension is 1×M l Is used for the feature vector of (a),M l representing the length of the feature vector,lthe value ranges from 1 toLThe method comprises the steps of carrying out a first treatment on the surface of the The number of the real fault characteristics and the counter-fact fault characteristics generated in the same layer network of the discriminator is the same;
the loss between the counter fact feature vector generated by the counter fact fault data in the layers of the network of the discriminator and the real feature vector generated by the target working condition fault data in the layers of the network of the discriminator is called layer contrast loss, and the layer contrast loss is expressed as follows:
wherein,L con l h,, representing the counterfactual fault data at the arbiterlThe inverse fact feature vector and the target working condition fault data generated in the layer network are in the first of the discriminatorslThe third layer is a layer network of the true feature vectorshLoss of individual layer contrast; />Representing the counterfactual fault data at the arbiterlGenerated in a layer networkhA counter-fact feature vector is used to determine,hthe value ranges from 1 toN l ，N l Representation of the discriminantlThe total number of the counter fact feature vectors or the real feature vectors generated in the layer network; />Representing the counterfactual fault data at the arbiterlGenerated in a layer network except->A feature vector set composed of all other inverse feature vectors except for the others; />Representation->The number of the middle-inverse fact feature vectors; />Representation->The first of (3)rA counter-fact feature vector is used to determine,rthe value range is 1 to->；/>Representation->The first of (3)fA plurality of inverse fact feature vectors; />Indicating the target working condition fault data in the first discriminatorlFeature vector set composed of all true feature vectors generated in layer networkCombining; />Representation->The first of (3)rTrue feature vectors;ωrepresenting the super-parameters;
the contrast losses of all layers in the discriminator are added to form the total contrast loss of the discriminator, and the total contrast loss is expressed as follows:
wherein (1)>Representing the total contrast loss value of the discriminator;Lindicating the total number of layers of the network in the arbiter.
3. The method for generating the anti-facts fault data based on causal intervention according to claim 2, wherein the anti-facts fault characteristics output by different layers of networks are selected from the generator for patch comparison, and a multi-layer patch comparison loss is obtained, and the comparison process is as follows:
S2A1, in the process of processing original working condition fault data by the generator, obtaining the counterfactual patches output by each layer of network from the generator, and inputting the counterfactual patches into the multi-layer perceptron network one by one, wherein the multi-layer perceptron network outputs the counterfactual patch characteristics corresponding to each layer of network, and the counterfactual patch characteristics are expressed as follows:
wherein (1)>Representation generatorGFirst, thebLayer network ofb i The true patches are inverse fact patches generated by the generating function;MLPrepresentation ofA multi-layer perceptron; />Representation->After the multi-layer perceptron is input, the corresponding output counterfactual patch features;
S2A2, calculating losses among the anti-fact patch features in the generator layer networks, the positive sample patch features in the anti-fact patch features and the negative sample patch features of the anti-fact patch features to obtain anti-fact patch losses of the anti-fact fault features, wherein the anti-fact patch losses are expressed as follows:
wherein,the counterfactual patch loss of original working condition fault data is represented; />Representation->Corresponding negative example patch feature, +.>Representation->Corresponding positive sample patch features;Orepresenting a loss function;Brepresenting the total number of layers of the network in the generator;b I the first of the representation generatorsbThe total number of counterfactual patch features in the layer network;
S2A3, inputting target working condition fault data into a generation countermeasure network, and processing the target working condition fault data according to the processing process of the original working condition fault data in the generator so as to obtain a corresponding real patch; processing the real patch according to the same processing mode as the step S2A1 and the step S2A2 to obtain the counterfactual patch loss corresponding to the target working condition fault data;
S2A4, combining the counterfactual patch loss of the original working condition fault data and the counterfactual patch loss of the target working condition fault data to obtain a multi-layer patch loss of the generator, wherein the multi-layer patch loss is expressed as follows:
wherein,a multi-layer patch loss representing a generator; />Representation->Weight parameters of (2); />The counterfactual patch loss of the target working condition fault data is represented; />Representation ofWeight parameters of (c).
4. A causal intervention based anti-facts fault data generation method according to claim 3, characterized in that the error optimisation is the simultaneous solution of the minimum of the counterdamage function in the CycleGAN network, generatorGMinimum and generator of causal relation losses of (a)GMulti-layer patch loss minima and arbiterDBy repeating the error optimization process and reaching a set stop condition, a new product can be producedAnd forming an optimal generator, and obtaining a mapping of the real fault characteristics in the original working condition fault data to the real fault characteristics in the target working condition fault data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410005964.9A CN117520905B (en) | 2024-01-03 | 2024-01-03 | Anti-fact fault data generation method based on causal intervention |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410005964.9A CN117520905B (en) | 2024-01-03 | 2024-01-03 | Anti-fact fault data generation method based on causal intervention |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117520905A CN117520905A (en) | 2024-02-06 |
| CN117520905B true CN117520905B (en) | 2024-03-22 |
Family
ID=89766778
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410005964.9A Active CN117520905B (en) | 2024-01-03 | 2024-01-03 | Anti-fact fault data generation method based on causal intervention |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117520905B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113824725A (en) * | 2021-09-24 | 2021-12-21 | 中国人民解放军国防科技大学 | Network safety monitoring analysis method and system based on causal machine learning |
| CN114897140A (en) * | 2022-05-09 | 2022-08-12 | 哈尔滨工业大学 | Counterfactual generation method based on causal intervention |
| KR20220139590A (en) * | 2021-04-08 | 2022-10-17 | 고려대학교 산학협력단 | Method and apparatus of generating a Counterfactual Map to explain the decision of Classifier |
| CN116108755A (en) * | 2023-03-09 | 2023-05-12 | 合肥工业大学 | Anti-fact confidence data generation method based on fault dictionary |
| CN116956005A (en) * | 2022-11-30 | 2023-10-27 | 腾讯科技（深圳）有限公司 | Training method, device, equipment, storage medium and product of data analysis model |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP4244767A1 (en) * | 2020-11-16 | 2023-09-20 | Umnai Limited | Method for an explainable autoencoder and an explainable generative adversarial network |
| US20220215243A1 (en) * | 2021-01-05 | 2022-07-07 | Capital One Services, Llc | Risk-Reliability Framework for Evaluating Synthetic Data Models |
-
2024
- 2024-01-03 CN CN202410005964.9A patent/CN117520905B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20220139590A (en) * | 2021-04-08 | 2022-10-17 | 고려대학교 산학협력단 | Method and apparatus of generating a Counterfactual Map to explain the decision of Classifier |
| CN113824725A (en) * | 2021-09-24 | 2021-12-21 | 中国人民解放军国防科技大学 | Network safety monitoring analysis method and system based on causal machine learning |
| CN114897140A (en) * | 2022-05-09 | 2022-08-12 | 哈尔滨工业大学 | Counterfactual generation method based on causal intervention |
| CN116956005A (en) * | 2022-11-30 | 2023-10-27 | 腾讯科技（深圳）有限公司 | Training method, device, equipment, storage medium and product of data analysis model |
| CN116108755A (en) * | 2023-03-09 | 2023-05-12 | 合肥工业大学 | Anti-fact confidence data generation method based on fault dictionary |
Non-Patent Citations (3)
| Title |
|---|
| "Counterfactual Faithful Data Generation Based on Disentangled Representation for Compound Fault Diagnosis of Rolling Bearings";Xu Ding.et al;《IEEE》;20230313;期刊摘要，1-3节 * |
| "COUNTERFACTUAL GENERATIVE NETWORKS";Axel Sauer.et al;《arXiv:2101.06046v1》;20210115;期刊摘要，1-3节 * |
| "半监督阶梯网络和GAN在滚动轴承故障诊断的应用";丁煦等;《机械设计与制造》;20220531(第5期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117520905A (en) | 2024-02-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhu et al. | A review of the application of deep learning in intelligent fault diagnosis of rotating machinery | |
| Jiang et al. | A GAN-based anomaly detection approach for imbalanced industrial time series | |
| Chen et al. | Deep transfer learning for bearing fault diagnosis: A systematic review since 2016 | |
| Zhao et al. | Deep multi-scale convolutional transfer learning network: A novel method for intelligent fault diagnosis of rolling bearings under variable working conditions and domains | |
| Huang et al. | A robust weight-shared capsule network for intelligent machinery fault diagnosis | |
| Huang et al. | Deep ensemble capsule network for intelligent compound fault diagnosis using multisensory data | |
| He et al. | Deep variational autoencoder classifier for intelligent fault diagnosis adaptive to unseen fault categories | |
| Ren et al. | A systematic review on imbalanced learning methods in intelligent fault diagnosis | |
| Ding et al. | Transfer learning for remaining useful life prediction across operating conditions based on multisource domain adaptation | |
| CN109655259A (en) | Combined failure diagnostic method and device based on depth decoupling convolutional neural networks | |
| Huang et al. | Memory residual regression autoencoder for bearing fault detection | |
| Li et al. | Fusing convolutional generative adversarial encoders for 3D printer fault detection with only normal condition signals | |
| Chen et al. | Deep residual network for identifying bearing fault location and fault severity concurrently | |
| Huang et al. | A transferable capsule network for decoupling compound fault of machinery | |
| Shi et al. | DecouplingNet: A stable knowledge distillation decoupling net for fault detection of rotating machines under varying speeds | |
| Wang et al. | Deep multiadversarial conditional domain adaptation networks for fault diagnostics of industrial equipment | |
| Wang et al. | A balanced adversarial domain adaptation method for partial transfer intelligent fault diagnosis | |
| Chen et al. | An Intelligent Fault Diagnostic Method Based on 2D-gcForest and L ${} _ {\text {2, p}} $-PCA Under Different Data Distributions | |
| Zhuang et al. | Fault diagnosis of bearings using a two-stage transfer alignment approach with semantic consistency and entropy loss | |
| Zhang et al. | CBAM-CRLSGAN: A novel fault diagnosis method for planetary transmission systems under small samples scenarios | |
| CN117520905B (en) | Anti-fact fault data generation method based on causal intervention | |
| CN112163474A (en) | Intelligent gearbox diagnosis method based on model fusion | |
| Hu et al. | An improved metalearning framework to optimize bearing fault diagnosis under data imbalance | |
| Zhang et al. | A flexible monitoring framework via dynamic-multilayer graph convolution network | |
| Feng et al. | Cross-Sensor Correlative Feature Learning and Fusion for Intelligent Fault Diagnosis |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |