Adaptive fault diagnosis method for monitoring data field of digital twin-driven mine hoist

Abstract

The invention discloses a digital twin-driven mine hoist monitoring data field adaptive fault diagnosis method, and the method comprises the steps: building four groups of physical entity models on a test bench, and obtaining a measurement target field data set; respectively establishing an outer ring fault digital twinning model, an inner ring fault digital twinning model, a roller fault digital twinning model and a health model digital twinning model, and obtaining corresponding simulated fault data sets; constructing a fault diagnosis model based on domain adaptation; constructing a domain adversarial neural network; sending the simulation source sample and the actually measured target sample into a domain adversarial neural network at the same time, reducing the feature distribution difference between the simulated source domain data and the measured target domain data by using the domain adversarial neural network, and obtaining a fault diagnosis model based on domain adaptation; and performing real-time fault diagnosis by using the fault diagnosis model based on domain adaptation, and outputting a fault diagnosis result. According to the method, the problem of inconsistent distribution of the test data and the training data can be effectively solved, and the accuracy and timeliness of fault diagnosis can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent fault diagnosis, and specifically is a digital twin driven mine hoist monitoring data domain adaptive fault diagnosis method. Background Art

[0002] As one of the core equipment of mining, mine hoists undertake important tasks such as transporting coal, mining equipment and personnel. The stability and reliability of their performance are directly related to the safe production and economic benefits of coal mines. Bearings are the core components that support the entire hoist, and their health status directly affects the overall performance of the hoist. Therefore, the research on hoist bearing fault diagnosis has important theoretical significance and practical value.

[0003] Traditional fault diagnosis methods mostly rely on manual experience and regular maintenance, and are difficult to adapt to the rapidly changing production environment. With the development of the big data era, algorithms such as machine learning and deep learning have gradually been widely used in the field of hoist fault diagnosis due to their unique advantages. Compared with traditional diagnostic methods, fault diagnosis technology based on big data can greatly improve the accuracy of fault diagnosis and significantly reduce the risk of false alarms and missed alarms. Although some deep learning methods can achieve excellent performance in the field of fault diagnosis, they usually have serious data dependence and require large-scale fault data sets to be obtained in advance as a training basis. However, in many key industrial scenarios, collecting large-scale training data is very expensive and difficult, which weakens the universal applicability of these deep learning technologies. With the rapid development of information technology, digital twin technology has gradually emerged. It can build virtual models through physical entities and generate a large amount of high-quality and diverse simulation fault data, thus providing an effective solution to the problem of insufficient hoist fault data.

[0004] In addition, the operating environment of mine hoists is complex. Affected by factors such as terrain, equipment installation location, and operating condition fluctuations, monitoring data often has strong time-varying and nonlinear characteristics. However, the test data and training data of most fault diagnosis methods currently come from the same data distribution. In actual situations, the inconsistent distribution of training data and test data caused by the variability of monitoring data will affect the diagnostic performance. Domain adaptation technology has great potential in solving the field of fault diagnosis under data distribution differences. It can effectively cope with the differences in data distribution under different environments and working conditions, and improve the generalization ability and diagnostic accuracy of the model in new environments. Summary of the invention

[0005] In response to the problems existing in the above-mentioned prior art, the present invention provides a digital twin-driven mine hoist monitoring data domain adaptive fault diagnosis method. The method can provide a large amount of high-quality hoist fault diagnosis data through digital twin technology, and can obtain a fault diagnosis model based on domain adaptation. It can effectively solve the problem of inconsistent distribution of test data and training data, and can adapt to data changes under different working conditions, thereby improving the accuracy, timeliness and reliability of diagnosis.

[0006] In order to achieve the above object, the present invention provides a digital twin driven mine hoist monitoring data domain adaptive fault diagnosis method, comprising the following steps:

[0007] Step 1: Establish a real space physical entity model and collect measurement target domain data;

[0008] S11: Build four groups of physical entity models on the test bench; each group of physical entity models includes a bearing seat, a rolling bearing, a drive motor, and an accelerometer, wherein the bearing seat and the drive motor are fixedly installed on the test bench, the rolling bearing is rotatably installed in the bearing seat, and the center of its inner ring is connected to the output shaft of the drive motor; the accelerometer is installed in the vertical direction of the bearing seat to collect vibration signals in the vertical direction;

[0009] S12: using one group of physical entity models as a healthy model; for the other three groups of physical entity models, manually setting faults, wherein an outer ring fault is set in the first group of physical entity models to form an outer ring fault model, an inner ring fault is set in the second group of physical entity models to form an inner ring fault model, and a ball fault is set in the third group of physical entity models to form a ball fault model;

[0010] S13: for the outer ring fault model, the inner ring fault model, the ball fault model and the healthy model, respectively, under the conditions that the speed of the driving motor is 1200 rpm and 1800 rpm, a real outer ring fault data set, an inner ring fault data set, a ball fault data set and a healthy condition data set are collected;

[0011] Step 2: Build a universal high-fidelity digital twin model of rolling bearings and collect simulation data;

[0012] The outer ring fault digital twin model, the inner ring fault digital twin model, the roller fault digital twin model and the healthy model digital twin model are respectively established. The outer ring fault digital twin model, the inner ring fault digital twin model, the roller fault digital twin model and the healthy model digital twin model correspond to the outer ring fault model, the inner ring fault model, the ball fault model and the healthy model respectively; for each digital twin model, the inner ring and the outer ring have two degrees of freedom respectively, namely the horizontal translation displacement x i 、x o, vertical translation displacement y i ,y o , and the unit resonator has only one degree of freedom, namely the vertical translation displacement y r ;Use the universal high-fidelity digital twin model to simulate the outer ring fault, inner ring fault and roller fault respectively, and obtain the corresponding simulated fault data set;

[0013] The local defects in the digital twin model of the outer ring fault are modeled in the form of rectangular fragments. The specific process is as follows:

[0014] A1: According to formula (1), the circumferential span angle β of the outer ring defect is obtained o ;

[0015]

[0016] Where L0 is the width of the outer ring defect, and D0 represents the diameter of the outer ring;

[0017] A2: Obtain the position relationship between the jth roller and the defect according to formula (2);

[0018] |mod(θ j ,2π)-θ D |<β o (2);

[0019] In the formula, θ D is the angular position of the defect, θ j is the angular position of the roller;

[0020] A3: According to formula (3), the contact deformation loss δ of the roller when the roller passes through the defect position is obtained o ;

[0021]

[0022] Where d is the roller diameter;

[0023] A4: According to formula (4), the contact deformation δ between the roller and the raceway when the roller passes the defect position is obtained j :

[0024] δ j =(x i -x o )cosθ j +(y i -y o )sinθ j -γ-δ o (4);

[0025] Where, γ is the radial clearance;

[0026] A5: The vibration response of the bearing is solved by the fourth-order fixed-step Runge-Kutta method, and the simulated outer race fault data set is obtained as the simulated source domain data of the outer race fault. The calculation time step is 5×10 -5 s;

[0027] The local defects in the inner loop fault digital twin model are modeled in the form of rectangular fragments. The specific process is as follows:

[0028] B1: According to formula (5), the circumferential span angle β of the inner ring defect is obtained i ;

[0029]

[0030] Where, L i is the width of the inner ring defect, D i Indicates the diameter of the inner ring;

[0031] B2: According to formula (6), the position relationship between the defect and the ball when they are in contact during the rotation of the inner ring is obtained;

[0032] |mod(θ i ,2π)-mod(θ j ,2π)|<β i (6);

[0033] In the formula, θ i is the angular position of the inner ring defect, θ i =ω s ×t;

[0034] B3: According to formula (7), the contact deformation loss δ when the defect contacts the ball is obtained i ;

[0035]

[0036] B4: According to formula (8), the contact deformation δ between the ball and the raceway when the ball passes through the internal defect is obtained j ;

[0037] δ j =(x i -x o )cosθ j +(y i -y o )sinθ j -γ-δ i (8);

[0038] B5: The vibration response of the bearing is solved by the fourth-order fixed-step Runge-Kutta method, and the simulated inner race fault data set is obtained as the simulated source domain data of the inner race fault. The calculation time step is 5×10-5 s;

[0039] The local defects in the roller fault digital twin model are modeled in the form of rectangular fragments. The specific process is as follows:

[0040] C1: The circumferential span angle β of the roller defect is obtained according to formula (9): r ;

[0041]

[0042] Where, L r is the roller defect width;

[0043] C2: The position relationship between the roller defect and the raceway when in contact is obtained according to formula (10) and formula (11);

[0044]

[0045] In the formula, is the angular position of the roller defect, Where D is the pitch diameter;

[0046] C3: According to formula (12), the contact deformation loss δ when the defect contacts the raceway is obtained r ;

[0047]

[0048] C4: According to formula (13), the contact deformation δ between the roller and the raceway when the defect contacts the raceway is obtained j ;

[0049] δ j =(x i -x o )cosθ j +(y i -y o )sinθ j -γ-δ r (13);

[0050] C5: The vibration response of the bearing is solved by the fourth-order fixed-step Runge-Kutta method, and the simulated ball fault data set is obtained as the simulated source domain data of the ball fault. The calculation time step is 5×10 -5 s;

[0051] Step 3: Build a fault diagnosis model based on domain adaptation;

[0052] S31: Constructing domain adversarial neural networks;

[0053] Four residual blocks are sequentially arranged between the five convolutional combination layers, and are sequentially connected to form an input side network, wherein each convolutional combination layer is composed of a convolutional layer, a batch normalization layer, an LReLU activation function layer and a maximum pooling layer stacked in sequence, the convolutional layer kernel size in the five convolutional combination layers is 9, the output channels are 4, 8, 16, 32 and 64 respectively, the maximum pooling layer kernel size and step size in the five convolutional combination layers are both set to 2, each residual block is composed of a convolutional layer, an LReLU layer and a batch normalization layer stacked in sequence, and the kernel size of the convolutional layer in each residual block is set to 3; an output side network is formed by two sequentially connected fully connected combination layers, and the input end of the output side network is connected to the output end of the input side network to form a feature extractor F, wherein each fully connected combination layer is composed of a fully connected layer, a batch normalization layer, an LReLU activation function layer and a Dropout layer stacked in sequence, and the fully connected layer output channels in the two fully connected combination layers are 300 and 100 respectively;

[0054] On the output side of the feature extractor F, the classifier C and the domain discriminator D are respectively connected to form a domain adversarial neural network, where the classifier C consists of a fully connected layer and the domain discriminator D is a domain discriminator based on adversarial learning;

[0055] S32: Simultaneously send the simulated source samples and the measured target samples into the domain adversarial neural network, and use the domain adversarial neural network to reduce the feature distribution difference between the simulated source domain data and the measured target domain data; first, a one-dimensional convolutional residual network is used through the feature extractor F to capture the local and detail features in the one-dimensional sensor data, and then the features are input into the classifier C to obtain the predicted labels of the simulated source domain data and the measured target domain data. At the same time, the domain discriminator D is used to learn the invariant feature representations of the simulated source domain data and the measured target domain data, and finally, a fault diagnosis model based on domain adaptation is obtained;

[0056] D1: Classifier C uses the cross entropy loss function in formula (14) to learn the discriminative features for classification and obtains the predicted output according to formula (15)

[0057]

[0058] Where n s is the sample size, x s and s are simulated source domain samples and labels, is the cross entropy loss, Among them, y is the label of the input sample x, is the predicted label, K is the number of health conditions; x iInput sample; F is defined as the shared feature space of the simulated source domain and the measured target domain, which is used to describe x i the probability assigned to each health state;

[0059] D2: A partial domain adaptation method is adopted to introduce a domain discriminator D based on adversarial learning to learn feature representations that are invariant to the simulation source domain and the measurement target domain, thereby achieving efficient cross-domain fault diagnosis;

[0060] D21: Order Represents the simulated source domain data, where n s is the number of samples in the simulated source domain, x s and s are the samples and labels of the simulated source domain respectively; at the same time, the label set of the simulated source domain is defined as C s ;make Represents the measured target domain data, where n t is the number of samples measuring the target domain, x t is the sample of the target domain; at the same time, the label set of the target domain is defined as C t ; Among them, |C s |>|C t |,C=C s ∩C t represents the shared label space, i.e., the categories that exist in both the simulation source domain and the measurement target domain; Represents the private class of the simulated source domain, also known as the exception class;

[0061] D22: Filter outlier conditions and relevant condition samples from the simulated source domain during domain adaptation, that is, introduce a selective adversarial network to achieve partial domain adaptation; split the discriminator D into |C s |k-th domain discriminator D divided by category k , to match the non-identical health state space C s ≠C t The simulated source domain data and the measured target domain data are as follows, where k represents the health status of the simulated source domain, i.e., k = 1,…,|C s |; Each D k Perform domain adaptation on samples with healthy state k in the simulated source domain and measured target domain, using the probability output in formula (15) To transform each input sample x i Arrange to the corresponding D k ;

[0062]

[0063] Where, d i Indicates the domain label;

[0064] D23: Introducing a gradient reversal layer into the shared feature layer for adversarial learning;

[0065] D24: Introduce a category-level weighting strategy to reduce the introduction of noise and improve the positive transfer effect by reducing the weight of the outlier health state domain discriminator in the simulated source domain, and obtain the noise loss according to formula (16):

[0066]

[0067] In the formula, represents the category-level weight of health state k;

[0068] D25: Introduce entropy minimization strategy to further enhance the separability between different health states and improve the prediction accuracy of target samples; this process minimizes the probability The prediction entropy loss To achieve this, the prediction entropy loss is obtained according to formula (17):

[0069]

[0070] Where H(·) is the conditional entropy loss;

[0071] D3: Combining the classifier loss and the domain discriminator loss, the total loss of the partial domain adaptation network is obtained according to formula (18) Finally, a fault diagnosis model based on domain adaptation is obtained;

[0072]

[0073] Where λ and β represent two trade-off parameters.

[0074] Step 4: Real-time fault diagnosis of rolling bearings;

[0075] The sensor group arranged on the bearing seat in the mine hoist is used to collect the vibration signal of the rolling bearing in real time, and the real-time vibration signal is input as input data into the fault diagnosis model based on domain adaptation. The fault diagnosis model based on domain adaptation is used to perform real-time fault diagnosis and output the fault diagnosis results.

[0076] As a preference, in step 1 S13, the sampling frequency is 20 khz.

[0077] As a preference, in S12 of step one, in the first group of physical entity models, the fault width of the outer ring is 1 mm, and the fault depth is 0.5 mm; in the second group of physical entity models, the fault width of the inner ring is 1 mm, and the fault depth is 0.5 mm; in the third group of physical entity models, the fault width of the ball is 1 mm, and the fault depth is 0.5 mm.

[0078] As a preferred embodiment, in the feature capture process of the feature extractor F in S32 of step three, in the convolution combination layer, the input bearing data is firstly convolved to extract local features in the input bearing data, and then the data is input into the batch normalization layer and the LReLU activation function layer to capture richer feature information, and then the data is input into the maximum pooling layer for maximum pooling operation to compress the input features in the spatial dimension and select the maximum value from each region as the output; in the residual block, the signal directly input after being processed by the previous convolution combination layer is processed to form a direct output signal, and then the direct output signal is added to the direct input signal after being processed by the previous convolution combination layer, and then output to the next convolution combination layer to realize information jump connection and capture more complex features; in the fully connected combination layer, the high-level feature map is flattened into a one-dimensional high-level feature vector by using the fully connected layer, and the generalization performance is enhanced by using the Dropout layer; in the classifier C, the fully connected layer is used to integrate all input features and extract global information.

[0079] As a preferred embodiment, in step 1 S11, the bearing seat and rolling bearing used to build the physical entity model are of the same model as the bearing seat and rolling bearing used in the mine hoist.

[0080] The present invention can effectively solve the thorny problem of insufficient training data for fault diagnosis methods by establishing a universal high-fidelity digital twin model of rolling bearings. Since the fault diagnosis method of hoist rolling bearings usually requires a large number of fault samples, it is very difficult to obtain such a data set in the industrial scene of the real world, and a large amount of time and manpower costs are required. The universal high-fidelity digital twin model can well simulate various operating conditions and fault failure modes of rolling bearings, so that sufficient simulated fault data sets can be easily obtained under low-cost conditions. Through four groups of physical entity models, one group of physical entity models is used as a healthy model, and the other three groups of physical entity models are respectively set with outer ring faults, inner ring faults and ball faults, and then based on obtaining measurement target domain data. Since the data sets used are respectively simulated source domain data and real measurement target domain data, there must be a large difference in the distribution of data in different domains. The present invention can effectively reduce the difference in feature distribution between simulated source domain data and measured target domain data by proposing a domain adversarial neural network. At the same time, since the fault type of the monitoring data collected in the actual industrial scene is usually unknown, that is, the fault category of the measured data is usually part of the fault category in the simulation data. In this case, achieving global domain adaptation of the source domain distribution and the target domain distribution will cause negative migration of the model due to samples of outlier conditions in the source domain. To this end, the present invention filters out samples of outlier conditions and related conditions from the source domain in the domain adversarial neural network, that is, introduces a selective adversarial network to achieve partial domain adaptation, and introduces a domain discriminator D based on adversarial learning to learn to simulate the source domain and measure the invariant feature representation of the target domain, so as to achieve efficient cross-domain fault diagnosis. By introducing a gradient reversal layer in the shared feature layer for adversarial learning, it is convenient to use cross entropy loss to achieve domain invariant learning, and help the model learn the common features between the source domain and the target domain to effectively suppress irrelevant features. By introducing a category-level weighting strategy, the impact of outlier samples can be further reduced and the connection between shared health states can be improved. By introducing an entropy minimization strategy, the separability between different health states can be further enhanced, and the prediction accuracy of target samples can be effectively improved.

[0081] The present invention provides a large amount of high-quality elevator fault diagnosis data through digital twin technology, and develops a fault diagnosis model based on domain adaptation, which effectively solves the problem of inconsistent distribution of test data and training data to adapt to data changes under different working conditions, thereby improving the accuracy, timeliness and reliability of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 It is the overall structure diagram of the domain adversarial neural network in the present invention;

[0083] Figure 2 is a schematic diagram of local defects of the digital twin model in the present invention;

[0084] Figure 3 This is the structural diagram of the feature extractor F of the present invention. DETAILED DESCRIPTION

[0085] The present invention will be further described below in conjunction with the accompanying drawings.

[0086] like Figures 1 to 3 As shown, the present invention provides a digital twin driven mine hoist monitoring data domain adaptive fault diagnosis method, comprising the following steps:

[0087] Step 1: Establish a real space physical entity model and collect measurement target domain data;

[0088] S11: In order to realize the digital twin technology, physical entities in real space are first required. For this purpose, four sets of physical entity models are built on the test bench; each set of physical entity models includes a bearing seat, a rolling bearing, a drive motor, and an accelerometer, wherein the bearing seat and the drive motor are fixedly installed on the test bench, the rolling bearing is rotatably installed in the bearing seat, and the center of its inner ring is connected to the output shaft of the drive motor; the accelerometer is installed in the vertical direction of the bearing seat to collect vibration signals in the vertical direction;

[0089] S12: using one group of physical entity models as a healthy model; for the other three groups of physical entity models, manually setting faults, wherein an outer ring fault (OF) is set in the first group of physical entity models to form an outer ring fault model, an inner ring fault (IF) is set in the second group of physical entity models to form an inner ring fault model, and a ball bearing fault (BF) is set in the third group of physical entity models to form a ball bearing fault model;

[0090] S13: for the outer ring fault model, the inner ring fault model, the ball fault model and the healthy model, respectively, under the conditions that the speed of the driving motor is 1200 rpm and 1800 rpm, a real outer ring fault data set, an inner ring fault data set, a ball fault data set and a healthy condition data set are collected;

[0091] Step 2: Build a universal high-fidelity digital twin model of rolling bearings and collect simulation data;

[0092] The fault diagnosis method of the hoist usually requires a large number of fault samples. However, in the real-world industrial scenario, it is very difficult to obtain such a data set. In order to solve the thorny problem of insufficient training data for the fault diagnosis method, it is necessary to establish a universal high-fidelity digital twin model of the rolling bearing. The universal high-fidelity digital twin model of the rolling bearing can well simulate various operating conditions and failure modes of the rolling bearing, so as to easily obtain a large number of simulated fault data sets.

[0093] The outer ring fault digital twin model, the inner ring fault digital twin model, the roller fault digital twin model and the healthy model digital twin model are respectively established. The outer ring fault digital twin model, the inner ring fault digital twin model, the roller fault digital twin model and the healthy model digital twin model correspond to the outer ring fault model, the inner ring fault model, the ball fault model and the healthy model respectively; for each digital twin model, the inner ring and the outer ring have two degrees of freedom respectively, namely the horizontal translation displacement x i 、x o , vertical translation displacement y i ,y o , and the unit resonator has only one degree of freedom, namely the vertical translation displacement y r ;Use the universal high-fidelity digital twin model to simulate the outer ring fault, inner ring fault and roller fault respectively, and obtain the corresponding simulated fault data set;

[0094] The local defects in the digital twin model of the outer ring fault are modeled in the form of rectangular fragments. The specific process is as follows:

[0095] A1: According to formula (1), the circumferential span angle β of the outer ring defect is obtained o ;

[0096]

[0097] Where L0 is the width of the outer ring defect, and D0 represents the diameter of the outer ring;

[0098] A2: Obtain the position relationship between the jth roller and the defect according to formula (2);

[0099] |mod(θ j ,2π)-θ D |<β o (2);

[0100] In the formula, θ D is the angular position of the defect, θ j is the angular position of the roller;

[0101] A3: According to formula (3), the contact deformation loss δ of the roller when the roller passes the defect position is obtained o ;

[0102]

[0103] Where d is the roller diameter;

[0104] A4: According to formula (4), the contact deformation δ between the roller and the raceway when the roller passes the defect position is obtained j :

[0105] δ j =(xi -x o )cosθ j +(y i -y o )sinθ j -γ-δ o (4);

[0106] Where, γ is the radial clearance;

[0107] A5: The vibration response of the bearing is solved by the fourth-order fixed-step Runge-Kutta method, and the simulated outer race fault data set is obtained as the simulated source domain data of the outer race fault. The calculation time step is 5×10 -5 s;

[0108] The local defects in the inner loop fault digital twin model are modeled in the form of rectangular fragments. The specific process is as follows:

[0109] B1: According to formula (5), the circumferential span angle β of the inner ring defect is obtained i ;

[0110]

[0111] Where, L i is the width of the inner ring defect, D i Indicates the diameter of the inner ring;

[0112] B2: According to formula (6), the position relationship between the defect and the ball when they are in contact during the rotation of the inner ring is obtained;

[0113] |mod(θ i ,2π)-mod(θ j ,2π)|<β i (6);

[0114] In the formula, θ i is the angular position of the inner ring defect, θ i =ω s ×t;

[0115] B3: According to formula (7), the contact deformation loss δ when the defect contacts the ball is obtained i ;

[0116]

[0117] B4: According to formula (8), the contact deformation δ between the ball and the raceway when the ball passes through the internal defect is obtained j ;

[0118] δ j =(x i -x o )cosθj +(y i -y o )sinθ j -γ-δ i (8);

[0119] B5: The vibration response of the bearing is solved by the fourth-order fixed-step Runge-Kutta method, and the simulated inner race fault data set is obtained as the simulated source domain data of the inner race fault. The calculation time step is 5×10 -5 s;

[0120] The local defects in the roller fault digital twin model are modeled in the form of rectangular fragments. The specific process is as follows:

[0121] C1: The circumferential span angle β of the roller defect is obtained according to formula (9): r ;

[0122]

[0123] Where, L r is the roller defect width;

[0124] C2: The position relationship between the roller defect and the raceway when in contact is obtained according to formula (10) and formula (11);

[0125]

[0126] In the formula, is the angular position of the roller defect, Where D is the pitch diameter;

[0127] C3: According to formula (12), the contact deformation loss δ when the defect contacts the raceway is obtained r ;

[0128]

[0129] C4: According to formula (13), the contact deformation δ between the roller and the raceway when the defect contacts the raceway is obtained j ;

[0130] δ j =(x i -x o )cosθ j +(y i -y o )sinθ j -γ-δ r (13);

[0131] C5: The vibration response of the bearing is solved by the fourth-order fixed-step Runge-Kutta method, and the simulated ball fault data set is obtained as the simulated source domain data of the ball fault. The calculation time step is 5×10 -5 s;

[0132] Step 3: Build a fault diagnosis model based on domain adaptation;

[0133] S31: Constructing domain adversarial neural networks;

[0134] Four residual blocks are set in sequence between the five convolutional combination layers and connected in sequence to form an input side network. Each convolutional combination layer consists of a convolutional layer, a batch normalization layer, and a Normalization), an LReLU activation function (LreLU) layer and a maximum pooling layer are stacked in sequence. The kernel size of the convolution layer in the five convolution combination layers is 9, and the output channels are 4, 8, 16, 32 and 64 respectively. The kernel size and stride of the maximum pooling layer in the five convolution combination layers are both set to 2. Each residual block is composed of a convolution layer, an LReLU layer and a batch normalization layer stacked in sequence. The kernel size of the convolution layer in each residual block is set to 3; the input side network is used to extract high-level feature maps from the input samples; the output side network is formed by using two fully connected combination layers connected in sequence, and the input end of the output side network is connected to the output end of the input side network to form a feature extractor F, wherein each fully connected combination layer is composed of a fully connected layer, a batch normalization layer, an LReLU activation function layer and a Dropout layer stacked in sequence, and the output channels of the fully connected layers in the two fully connected combination layers are 300 and 100 respectively; the output side network is used to flatten the high-level feature map into a one-dimensional feature vector;

[0135] The classifier C and the domain discriminator D are respectively connected on the output side of the feature extractor F to form a domain adversarial neural network, wherein the classifier C consists of a fully connected layer and is used to realize the health status identification of the rolling bearing; the domain discriminator D is a domain discriminator based on adversarial learning;

[0136] S32: Simultaneously send the simulated source samples and the measured target samples into the domain adversarial neural network, and use the domain adversarial neural network to reduce the feature distribution difference between the simulated source domain data and the measured target domain data; first, a one-dimensional convolutional residual network is used through the feature extractor F to capture the local and detail features in the one-dimensional sensor data, and then the features are input into the classifier C to obtain the predicted labels of the simulated source domain data and the measured target domain data. At the same time, the domain discriminator D is used to learn the invariant feature representations of the simulated source domain data and the measured target domain data, and finally, a fault diagnosis model based on domain adaptation is obtained;

[0137] D1: In order to effectively learn the discriminant features, the classifier C uses the cross entropy loss function in formula (14) to learn the discriminant features for classification, thereby realizing the fault diagnosis function and obtaining the prediction output according to formula (15)

[0138]

[0139] Where n s is the sample size, x s and s are simulated source domain samples and labels, is the cross entropy loss, Among them, y is the label of the input sample x, is the predicted label, K is the number of health conditions; x i Input sample; F is defined as the shared feature space of the simulated source domain and the measured target domain, which is used to describe x i the probability assigned to each health state;

[0140] D2: In domain adversarial neural networks, the most important thing is to use adversarial learning strategies to achieve domain-invariant learning. Since the fault categories of measured data are usually part of the fault categories in simulation data, a partial domain adaptation method is adopted to introduce a domain discriminator D based on adversarial learning to learn feature representations that are invariant to the simulation source domain and the measurement target domain, thereby achieving efficient cross-domain fault diagnosis;

[0141] D21: Order Represents the simulated source domain data, where n s is the number of samples in the simulated source domain, x s and i are the samples and labels of the simulated source domain respectively; at the same time, the label set of the simulated source domain is defined as C s ;make Represents the measured target domain data, where n t is the number of samples measuring the target domain, x t is the sample of the target domain; at the same time, the label set of the target domain is defined as C t ; Among them, |C s |>|C t |,C=C s ∩C t represents the shared label space, i.e., the categories that exist in both the simulation source domain and the measurement target domain; Represents the private class of the simulated source domain, also known as the exception class;

[0142] D22: Filter outlier conditions and relevant condition samples from the simulated source domain during domain adaptation, that is, introduce a selective adversarial network to achieve partial domain adaptation; in order to match non-identical health state spaces C s ≠C t The simulated data and measured data of the discriminator D are split into |C s |k-th domain discriminator D divided by category k , to match the non-identical health state space C s ≠C t The simulated source domain data and the measured target domain data are as follows, where k represents the health status of the simulated source domain, i.e., k = 1,…,|C s |; Each D k Perform domain adaptation on samples with healthy state k in the simulated source domain and measured target domain, using the probability output in formula (15) To transform each input sample x i Arrange to the corresponding D k ;

[0143]

[0144] Where, d i Indicates the domain label;

[0145] D23: In order to achieve domain-invariant learning using cross entropy loss, a gradient reversal layer is introduced in the shared feature layer for adversarial learning. The gradient reversal layer is a special layer used for adversarial learning and domain adaptation. Its main function is to reverse the sign of the gradient during backpropagation, helping the model learn the common features between the source domain and the target domain and suppress irrelevant features.

[0146] D24: In order to further reduce the impact of outlier samples and improve the connection between shared health states, a category-level weighting strategy is introduced. In theory, the health state C of the target domain is t The domain discriminator of can promote the positive transfer of the model, while the outlier health state in the source domain The domain discriminator of usually introduces noise and weakens the positive transfer of the model. The introduction of noise is reduced and the positive transfer effect is improved by reducing the weight of the outlier health state domain discriminator in the simulated source domain, and the noise loss is obtained according to formula (16):

[0147]

[0148] In the formula, represents the category-level weight of health state k;

[0149] D25: Although L DIt can selectively transfer relevant diagnostic knowledge by reducing the negative impact of outlier health states, but it is highly dependent on Therefore, the entropy minimization strategy is introduced to further enhance the separability between different health states and improve the prediction accuracy of the target sample; this process minimizes the probability The prediction entropy loss To achieve this, the prediction entropy loss is obtained according to formula (17):

[0150]

[0151] Where H(·) is the conditional entropy loss;

[0152] D3: Combining the classifier loss and the domain discriminator loss, the total loss of the partial domain adaptation network is obtained according to formula (18) Finally, a fault diagnosis model based on domain adaptation is obtained;

[0153]

[0154] Where λ and β represent two trade-off parameters.

[0155] Step 4: Real-time fault diagnosis of rolling bearings;

[0156] The sensor group arranged on the bearing seat in the mine hoist is used to collect the vibration signal of the rolling bearing in real time, and the real-time vibration signal is input as input data into the fault diagnosis model based on domain adaptation. The fault diagnosis model based on domain adaptation is used to perform real-time fault diagnosis and output the fault diagnosis results.

[0157] As a preference, in step 1 S13, the sampling frequency is 20 khz.

[0158] Preferably, in S1 of step one, in the first group of physical entity models, the width of the fault of the outer ring is 1 mm, and the depth of the fault is 0.5 mm; in the second group of physical entity models, the width of the fault of the inner ring is 1 mm, and the depth of the fault is 0.5 mm; in the third group of physical entity models, the width of the fault of the ball is 1 mm, and the depth of the fault is 0.5 mm.

[0159] As a preferred embodiment, in the feature capture process of the feature extractor F in S32 of step three, in the convolution combination layer, a convolution operation is first performed on the input bearing data. The local features in the input bearing data can be effectively extracted through the convolution layer, and then the data is input into the batch normalization layer and the LreLU (Leaky Rectified Linear Unit) activation function layer. The LReLU (Leaky Rectified Linear Unit) activation function is an improved version of the ReLU (Rectified Linear Unit) and is intended to solve the problem of "dead neurons" in the negative input area of ​​the ReLU. It is simple to calculate, easy to implement, and performs well during training, and can capture richer feature information. After that, the data is input into the maximum pooling layer for maximum pooling operation. The main function of this layer is to compress the input features in spatial dimension. The maximum pooling selects the maximum value from each region as the output, which helps to highlight the significant features of the input data, and can also speed up the training of the model and reduce overfitting. In the residual block, the signal directly input after the previous convolution combination layer is processed to form a direct output signal, and then the direct output signal is added to the direct input signal after the previous convolution combination layer, and then output to the next convolution combination layer to realize the jump connection of information and capture more complex features. This structure can alleviate the gradient vanishing and explosion problems in deep networks, so that information can be more effectively propagated in the network, thereby improving training efficiency and model performance. At the same time, the residual block allows a deeper network structure, so that the model can capture more complex features. In the fully connected combination layer, the fully connected layer is used to flatten the high-level feature map into a one-dimensional high-level feature vector, and the Dropout layer is used to enhance the generalization performance. The Dropout operation is used in the two fully connected layers to effectively prevent the occurrence of overfitting. In each training iteration, some neurons are temporarily closed with a certain probability so that they do not participate in the calculation in the current iteration, thereby simplifying the network structure and enhancing the generalization performance of the model. In classifier C, the role of the fully connected layer is to connect all neurons in the previous layer to each neuron in the current layer, and the fully connected layer is used to integrate all input features and extract global information.

[0160] As a preferred embodiment, in step 1 S11, the bearing seat and rolling bearing used to build the physical entity model are of the same model as the bearing seat and rolling bearing used in the mine hoist.

[0161] The present invention can effectively solve the thorny problem of insufficient training data for fault diagnosis methods by establishing a universal high-fidelity digital twin model of rolling bearings. Since the fault diagnosis method of hoist rolling bearings usually requires a large number of fault samples, it is very difficult to obtain such a data set in the industrial scene of the real world, and a large amount of time and manpower costs are required. The universal high-fidelity digital twin model can well simulate various operating conditions and fault failure modes of rolling bearings, so that sufficient simulated fault data sets can be easily obtained under low-cost conditions. Through four groups of physical entity models, one group of physical entity models is used as a healthy model, and the other three groups of physical entity models are respectively set with outer ring faults, inner ring faults and ball faults, and then based on obtaining measurement target domain data. Since the data sets used are respectively simulated source domain data and real measurement target domain data, there must be a large difference in the distribution of data in different domains. The present invention can effectively reduce the difference in feature distribution between simulated source domain data and measured target domain data by proposing a domain adversarial neural network. At the same time, since the fault type of the monitoring data collected in the actual industrial scene is usually unknown, that is, the fault category of the measured data is usually part of the fault category in the simulation data. In this case, achieving global domain adaptation of the source domain distribution and the target domain distribution will cause negative migration of the model due to samples of outlier conditions in the source domain. To this end, the present invention filters out samples of outlier conditions and related conditions from the source domain in the domain adversarial neural network, that is, introduces a selective adversarial network to achieve partial domain adaptation, and introduces a domain discriminator D based on adversarial learning to learn to simulate the source domain and measure the invariant feature representation of the target domain, so as to achieve efficient cross-domain fault diagnosis. By introducing a gradient reversal layer in the shared feature layer for adversarial learning, it is convenient to use cross entropy loss to achieve domain invariant learning, and help the model learn the common features between the source domain and the target domain to effectively suppress irrelevant features. By introducing a category-level weighting strategy, the impact of outlier samples can be further reduced and the connection between shared health states can be improved. By introducing an entropy minimization strategy, the separability between different health states can be further enhanced, and the prediction accuracy of target samples can be effectively improved.

[0162] The present invention provides a large amount of high-quality elevator fault diagnosis data through digital twin technology, and develops a fault diagnosis model based on domain adaptation, which effectively solves the problem of inconsistent distribution of test data and training data to adapt to data changes under different working conditions, thereby improving the accuracy, timeliness and reliability of diagnosis.

Claims

1. A digital twin-driven mine hoist monitoring data domain adaptive fault diagnosis method, characterized in that: The following steps are involved: Step 1: Establish a real space physical entity model and collect measurement target domain data; S11: Build four groups of physical entity models on the test bench; each group of physical entity models includes a bearing seat, a rolling bearing, a drive motor, and an accelerometer, wherein the bearing seat and the drive motor are fixedly installed on the test bench, the rolling bearing is rotatably installed in the bearing seat, and the center of its inner ring is connected to the output shaft of the drive motor; the accelerometer is installed in the vertical direction of the bearing seat to collect vibration signals in the vertical direction; S12: using one group of physical entity models as a healthy model; for the other three groups of physical entity models, manually setting faults, wherein an outer ring fault is set in the first group of physical entity models to form an outer ring fault model, an inner ring fault is set in the second group of physical entity models to form an inner ring fault model, and a ball fault is set in the third group of physical entity models to form a ball fault model; S13: for the outer ring fault model, the inner ring fault model, the ball fault model and the healthy model, respectively, under the conditions that the speed of the driving motor is 1200 rpm and 1800 rpm, a real outer ring fault data set, an inner ring fault data set, a ball fault data set and a healthy condition data set are collected; Step 2: Build a universal high-fidelity digital twin model of rolling bearings and collect simulation data; The outer ring fault digital twin model, the inner ring fault digital twin model, the roller fault digital twin model and the healthy model digital twin model are respectively established. The outer ring fault digital twin model, the inner ring fault digital twin model, the roller fault digital twin model and the healthy model digital twin model correspond to the outer ring fault model, the inner ring fault model, the ball fault model and the healthy model respectively; for each digital twin model, the inner ring and the outer ring have two degrees of freedom respectively, namely the horizontal translation displacement x i 、x o , vertical translation displacement y i ,y o , and the unit resonator has only one degree of freedom, namely the vertical translation displacement y r ;Use the universal high-fidelity digital twin model to simulate the outer ring fault, inner ring fault and roller fault respectively, and obtain the corresponding simulated fault data set; The local defects in the digital twin model of the outer ring fault are modeled in the form of rectangular fragments. The specific process is as follows: A1: According to formula (1), the circumferential span angle β of the outer ring defect is obtained o ; Where L0 is the width of the outer ring defect, and D0 represents the diameter of the outer ring; A2: Obtain the position relationship between the jth roller and the defect according to formula (2); |mod(θ j ,2π)-θ D |<β o (2); In the formula, θ D is the angular position of the defect, θ j is the angular position of the roller; A3: According to formula (3), the contact deformation loss δ of the roller when the roller passes through the defect position is obtained o ; Where d is the roller diameter; A4: According to formula (4), the contact deformation δ between the roller and the raceway when the roller passes the defect position is obtained j : d j =(x i -x o )cosθ j +(y i -y o )sinθ j -c-d o (4); Where, γ is the radial clearance; A5: The vibration response of the bearing is solved by the fourth-order fixed-step Runge-Kutta method, and the simulated outer race fault data set is obtained as the simulated source domain data of the outer race fault. The calculation time step is 5×10 -5 s; The local defects in the inner loop fault digital twin model are modeled in the form of rectangular fragments. The specific process is as follows: B1: According to formula (5), the circumferential span angle β of the inner ring defect is obtained i ; Where, L i is the width of the inner ring defect, D i Indicates the diameter of the inner ring; B2: According to formula (6), the position relationship between the defect and the ball when they are in contact during the rotation of the inner ring is obtained; |mod(θ i ,2π)-mod(θ j ,2ω)|<β i (6); In the formula, θ i is the angular position of the inner ring defect, θ i =ω s ×t; B3: According to formula (7), the contact deformation loss δ when the defect contacts the ball is obtained i ; B4: According to formula (8), the contact deformation δ between the ball and the raceway when the ball passes through the internal defect is obtained j ; d j =(x i -x o )cosθ j +(y i -y o )sinθ j -c-d i (8); B5: The vibration response of the bearing is solved by the fourth-order fixed-step Runge-Kutta method, and the simulated inner race fault data set is obtained as the simulated source domain data of the inner race fault. The calculation time step is 5×10 -5 s; The local defects in the roller fault digital twin model are modeled in the form of rectangular fragments. The specific process is as follows: C1: The circumferential span angle β of the roller defect is obtained according to formula (9): r ; Where, L r is the roller defect width; C2: The position relationship between the roller defect and the raceway when in contact is obtained according to formula (10) and formula (11); In the formula, is the angular position of the roller defect, Where D is the pitch diameter; C3: According to formula (12), the contact deformation loss v when the defect contacts the raceway is obtained r ; C4: According to formula (13), the contact deformation δ between the roller and the raceway when the defect contacts the raceway is obtained j ; d j =(x i -x o )cosθ j +(y i -y o )sinθ j -c-d r (13); C5: The vibration response of the bearing is solved by the fourth-order fixed-step Runge-Kutta method, and the simulated ball fault data set is obtained as the simulated source domain data of the ball fault. The calculation time step is 5×10 -5 s; Step 3: Build a fault diagnosis model based on domain adaptation; S31: Constructing domain adversarial neural networks; Four residual blocks are sequentially arranged between the five convolutional combination layers, and are sequentially connected to form an input side network, wherein each convolutional combination layer is composed of a convolutional layer, a batch normalization layer, an LReLU activation function layer and a maximum pooling layer stacked in sequence, the convolutional layer kernel size in the five convolutional combination layers is 9, the output channels are 4, 8, 16, 32 and 64 respectively, the maximum pooling layer kernel size and step size in the five convolutional combination layers are both set to 2, each residual block is composed of a convolutional layer, an LReLU layer and a batch normalization layer stacked in sequence, and the kernel size of the convolutional layer in each residual block is set to 3; an output side network is formed by two sequentially connected fully connected combination layers, and the input end of the output side network is connected to the output end of the input side network to form a feature extractor F, wherein each fully connected combination layer is composed of a fully connected layer, a batch normalization layer, an LReLU activation function layer and a Dropout layer stacked in sequence, and the fully connected layer output channels in the two fully connected combination layers are 300 and 100 respectively; On the output side of the feature extractor F, the classifier C and the domain discriminator D are respectively connected to form a domain adversarial neural network, where the classifier C consists of a fully connected layer and the domain discriminator D is a domain discriminator based on adversarial learning; S32: Simultaneously send the simulated source samples and the measured target samples into the domain adversarial neural network, and use the domain adversarial neural network to reduce the feature distribution difference between the simulated source domain data and the measured target domain data; first, a one-dimensional convolutional residual network is used through the feature extractor F to capture the local and detail features in the one-dimensional sensor data, and then the features are input into the classifier C to obtain the predicted labels of the simulated source domain data and the measured target domain data. At the same time, the domain discriminator D is used to learn the invariant feature representations of the simulated source domain data and the measured target domain data, and finally, a fault diagnosis model based on domain adaptation is obtained; D1: Classifier C uses the cross entropy loss function in formula (14) to learn the discriminative features for classification and obtains the predicted output according to formula (15) Where n s is the sample size, x s and s are simulated source domain samples and labels, is the cross entropy loss, Among them, y is the label of the input sample x, is the predicted label, K is the number of health conditions; x i Input sample; F is defined as the shared feature space of the simulated source domain and the measured target domain, which is used to describe x i the probability assigned to each health state; D2: A partial domain adaptation method is adopted to introduce a domain discriminator D based on adversarial learning to learn feature representations that are invariant to the simulation source domain and the measurement target domain, thereby achieving efficient cross-domain fault diagnosis; D21: Order Represents the simulated source domain data, where n s is the number of samples in the simulated source domain, x s and s are the samples and labels of the simulated source domain respectively; at the same time, the label set of the simulated source domain is defined as C s ;make Represents the measured target domain data, where n t is the number of samples measuring the target domain, x t is the sample of the target domain; at the same time, the label set of the target domain is defined as C t ; Among them, |C s |>|C t |,C=C s ∩c t represents the shared label space, i.e., the categories that exist in both the simulation source domain and the measurement target domain; Represents the private class of the simulated source domain, also known as the exception class; D22: Filter outlier conditions and relevant condition samples from the simulated source domain during domain adaptation, that is, introduce a selective adversarial network to achieve partial domain adaptation; split the discriminator D into |c s |k-th domain discriminator D divided by category k , to match the non-identical health state space C s ≠C t The simulated source domain data and the measured target domain data are as follows, where k represents the health status of the simulated source domain, i.e., k = 1,…,|C s |; Each D k Perform domain adaptation on samples with healthy state k in the simulated source domain and measured target domain, using the probability output in formula (15) To transform each input sample x i Arrange to the corresponding D k ; Where, d i Indicates the domain label; D23: Introducing a gradient reversal layer into the shared feature layer for adversarial learning; D24: Introduce a category-level weighting strategy to reduce the introduction of noise and improve the positive transfer effect by reducing the weight of the outlier health state domain discriminator in the simulated source domain, and obtain the noise loss according to formula (16): In the formula, represents the category-level weight of health state k; D25: Introduce entropy minimization strategy to further enhance the separability between different health states and improve the prediction accuracy of target samples; this process minimizes the probability The prediction entropy loss To achieve this, the prediction entropy loss is obtained according to formula (17): Where H(·) is the conditional entropy loss; D3: Combining the classifier loss and the domain discriminator loss, the total loss of the partial domain adaptation network is obtained according to formula (18) Finally, a fault diagnosis model based on domain adaptation is obtained; Where λ and β represent two trade-off parameters. Step 4: Real-time fault diagnosis of rolling bearings; The sensor group arranged on the bearing seat in the mine hoist is used to collect the vibration signal of the rolling bearing in real time, and the real-time vibration signal is input as input data into the fault diagnosis model based on domain adaptation. The fault diagnosis model based on domain adaptation is used to perform real-time fault diagnosis and output the fault diagnosis results.

2. According to claim 1, a digital twin driven mine hoist monitoring data domain adaptive fault diagnosis method is characterized in that: In step 1 S13, the sampling frequency is 20 khz.

3. The method for adaptive fault diagnosis of mine hoist monitoring data domain driven by digital twin according to claim 1 is characterized in that: In step 1 S12, in the first group of physical entity models, the width of the fault of the outer ring is 1 mm, and the fault depth is 0.5 mm; in the second group of physical entity models, the width of the fault of the inner ring is 1 mm, and the fault depth is 0.5 mm; in the third group of physical entity models, the width of the fault of the ball is 1 mm, and the fault depth is 0.5 mm.

4. The method for adaptive fault diagnosis of mine hoist monitoring data domain driven by digital twin according to claim 1 is characterized in that: In the feature capture process of the feature extractor F in S32 of step three, in the convolution combination layer, the input bearing data is firstly convolved to extract the local features in the input bearing data, and then the data is input into the batch normalization layer and the LReLU activation function layer to capture richer feature information, and then the data is input into the maximum pooling layer for maximum pooling operation to compress the input features in the spatial dimension and select the maximum value from each region as the output; in the residual block, the signal directly input after being processed by the previous convolution combination layer is processed to form a direct output signal, and then the direct output signal is added to the direct input signal after being processed by the previous convolution combination layer, and then output to the next convolution combination layer to realize the jump connection of information and capture more complex features; in the fully connected combination layer, the fully connected layer is used to flatten the high-level feature map into a one-dimensional high-level feature vector, and the Dropout layer is used to enhance the generalization performance; in the classifier C, the fully connected layer is used to integrate all input features and extract global information.

5. The method for adaptive fault diagnosis of mine hoist monitoring data domain driven by digital twin according to claim 1 is characterized in that: In step 1 S11, the bearing seat and rolling bearing used to build the physical entity model are the same as the bearing seat and rolling bearing used in the mine hoist.

Images