A method and system for fault diagnosis of AC variable frequency asynchronous motor in oil drilling rigs
By integrating multi-sensor information fusion and deep learning models, combined with improved evidence theory, efficient fault diagnosis of AC variable frequency asynchronous motors in oil drilling environments has been achieved, improving diagnostic accuracy and adaptability while reducing safety risks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHANGZHOU UNIV
- Filing Date
- 2023-06-16
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies cannot effectively extract fault characteristics of AC variable frequency asynchronous motors in oil drilling environments, resulting in low diagnostic accuracy and an inability to adapt to motor fault diagnosis in complex environments.
A multi-sensor state parameter observation and information fusion diagnostic mechanism is adopted, which combines CNN neural network and ResNet deep network. Through adaptive sparse modules and feature transfer model, decision-level fusion is performed using improved DS evidence theory to achieve fault identification.
It improves the accuracy of fault diagnosis for AC variable frequency asynchronous motors in oil drilling rigs, adapts to complex environments, reduces the misdiagnosis rate of motor faults, and reduces maintenance costs and safety risks.
Smart Images
Figure CN116861183B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of oil drilling and motor fault diagnosis technology, and in particular to a method and system for diagnosing faults in an AC variable frequency asynchronous motor of an oil drilling rig. Background Technology
[0002] As the core driving force in oil drilling operations, the reliable operation of AC variable frequency asynchronous motors directly affects the stability of production. The complex oil drilling environment, with its dynamic interactions and coupling effects of factors such as salt and alkali corrosion, high and low ambient temperatures, and sandstorms and snowstorms, influences the motor's condition and performance. Frequent relocation and disassembly of drilling rigs, and drilling jumps during operations can all cause misalignment between the motor and gearbox, leading to vibration, increased shaft temperature, accelerated shaft wear, uneven rotor air gap, and increased shaft current. Failure to detect these problems in time can result in motor bearing burnout and coupling damage. This not only increases maintenance costs but may also lead to major drilling safety accidents such as drilling stoppages, drill bit jams, and stuck drill bits.
[0003] Currently, many experts are conducting research on motor fault diagnosis, but the application scenarios in oil drilling have not been addressed. In the oil drilling environment, extracting different fault features for diagnosing motor faults is interactive and varied. Under the influence of internal and external factors such as power supply harmonics, high and low ambient temperatures, sudden load changes, and complex formations, existing methods for diagnosing motor faults by extracting single fault features are not entirely applicable and have low accuracy. Summary of the Invention
[0004] To address the shortcomings of existing methods, this invention establishes a multi-sensor state parameter observation and information fusion diagnostic mechanism that can adapt to different parameter feature extraction methods and fault identification methods, and constructs a fusion control model with different identification methods.
[0005] The technical solution adopted in this invention is: a fault diagnosis method for AC variable frequency asynchronous motors in oil drilling rigs, comprising the following steps:
[0006] Step 1: Collect vibration signals of the motor in the x, y, and z axes under normal and fault conditions, input the vibration signals into a CNN neural network, and output fault identification results based on the vibration signals.
[0007] Furthermore, the types of motor failures include: static eccentricity, dynamic eccentricity, bearing failure, and rotor bar breakage.
[0008] Furthermore, step one specifically includes:
[0009] The collected vibration signals are input into a one-dimensional CNN, and after two convolution and pooling operations, the high-dimensional and low-dimensional features of the signals are extracted.
[0010] An adaptive sparse module is designed and embedded in a CNN to dynamically generate attention weights for each feature vector output by the CNN front-end network, adaptively calculate the threshold of the attention weights, and filter redundant vectors in real time.
[0011] Furthermore, the adaptive sparse module includes:
[0012] A threshold estimation module is added to the attention mechanism. The attention layer uses the sigmoid function to obtain an initial threshold α, as shown in the formula:
[0013] α = sigmoid(W a z+b a (10)
[0014] Where z is the output of the attention layer, which is a one-dimensional vector; W a and b a These are the weights and biases, respectively.
[0015] The final threshold τ is constructed using the following formula:
[0016] τ=α·max(z) (11)
[0017] The masking function is constructed to filter the output of the attention layer, and the formula is as follows:
[0018]
[0019] Among them, ReLU is a commonly used activation function;
[0020] Finally, the filtered output and the output of the attention layer are recalculated to reconstruct the output of the attention layer. The output formula of the sparse attention module is:
[0021]
[0022] Where, N seg Number of sequences;
[0023] The feature vector output of the adaptive sparse module is divided into a training set and a test set. First, the training set is input into the Softmax layer for training. When the diagnostic accuracy converges, the training is complete. Then, the test set is input into the Softmax layer for testing, and the fault identification result is output.
[0024] Step 2: Collect the three-phase current signals of the motor under normal and fault conditions, input them into the ResNet deep network, and output the fault identification results based on the current signals.
[0025] Furthermore, step two specifically includes:
[0026] Step 21: Divide the acquired current signal into source domain samples, target domain training set samples, and target domain test set samples;
[0027] Step 22: Construct a ResNet deep network and initialize the network's weights and parameters;
[0028] Step 23: Input the source and target domain training data into the ResNet deep network to extract deep features from the source and target domains;
[0029] Step 24: Combine the feature transfer model, share the network parameters of the two domains, and simultaneously map the deep features extracted from the source domain and the target domain to the regenerating kernel Hilbert space (RKHS).
[0030] Step 25: After the fully connected layer of the ResNet deep network, JMMD is introduced to measure the distance between the feature spaces of the two domains. By minimizing the loss function, the difference in feature distribution between the source domain and the target domain is reduced, enabling the model to extract features with high similarity and update the ResNet deep network parameters during backpropagation.
[0031] Step 26: Input the target domain test set data into the updated ResNet deep network for feature extraction and output the fault identification results.
[0032] Furthermore, the JMMD loss function is calculated as follows:
[0033]
[0034] In the formula, H l Let |L| represent the l-th level RKHS, where |L| is the level number of the corresponding set. For the eigenmap of the tensor product in RKHS, z sl With z tl E represents the activation of the source and target domains in layer l, respectively. p E Q This is the mean distance between the vectors in the source and target domains.
[0035] Furthermore, embedding JMMD into the feature transfer network yields the following loss function:
[0036] ι=ι c +λι JMMD (D s D t (15)
[0037] In the formula, ι is the loss function of the entire feature transfer network, ι c Let represent the classification loss function of the source domain, and λ be the trade-off parameter for the overall network loss function; D s D t This represents the spatial distance between the source and target domains.
[0038] Step 3: Input the fault identification results based on current signals and the fault identification results based on vibration signals into the improved DS evidence theory model for decision-level fusion output.
[0039] Further improvements to the DS evidence theory model include:
[0040] Step 1: Calculate the cosine similarity between any two pieces of evidence;
[0041] Step 2: Cosine similarity between various pieces of evidence (sim) ij Form a similarity matrix S;
[0042]
[0043] Step 3: Calculate the evidence support by summing the similarity scores of all evidence-related items in the matrix.
[0044]
[0045] Define the average similarity of the evidence;
[0046]
[0047] Find the weighting coefficient β i ;
[0048]
[0049] Using weighting coefficient β i To refine the source of evidence; use β i The BPA of all evidence is weighted and averaged. Then, the weighted average evidence is combined n-1 times using the DS fusion rule to obtain the final fusion result.
[0050] A fault diagnosis system for an AC variable frequency asynchronous motor of an oil drilling rig includes: a signal acquisition module and a central processing module. The signal acquisition module acquires the operating data of the motor at the drilling site in real time, and acquires the vertical or horizontal vibration signal and the current signal of the motor through vibration sensors and current sensors, respectively. The acquired vibration and current signals are wirelessly transmitted to the central processing module after being conditioned by a conditioning circuit.
[0051] The central processing module includes a signal analysis unit and a fault diagnosis unit. The signal analysis unit uses CNN neural network and ResNet deep network to extract feature vectors from vibration and current signals, respectively, and performs fault identification.
[0052] The fault diagnosis unit uses an improved DS evidence theory to perform decision-level fusion of the two fault identification results to obtain the final fault diagnosis result.
[0053] Furthermore, the central processing module also includes an algorithm porting unit. The algorithm porting unit compiles the motor model, the CNN neural network and ResNet deep network model in the signal analysis unit, and the DS evidence theory model in the fault diagnosis unit to form a callable, interfaceable, and portable software package, which can be seamlessly integrated with the intelligent operation and maintenance platform for oil drilling equipment.
[0054] The beneficial effects of this invention are:
[0055] 1. Considering the coupled influence of natural factors in the drilling environment and motor operating parameters, a multi-sensor state parameter observation and information fusion diagnostic mechanism was established to improve the fault diagnosis accuracy of AC variable frequency asynchronous motors in oil drilling rigs;
[0056] 2. The feature vector of the vibration signal is extracted using CNN, an adaptive sparse module is designed to filter out redundant vectors, and the fault identification result based on the vibration signal is realized through the Softmax layer.
[0057] 3. The ResNet deep network model is used to extract deep features of the current signal. Combined with the feature transfer model, JMMD is introduced to adapt the feature space distribution. While finding domain-invariant features, the adaptability of the diagnostic model in the drilling environment is improved.
[0058] 4. A DS evidence fusion algorithm based on cosine similarity is adopted to perform decision-level fusion of the two identification results, realizing the final fault diagnosis, which is more suitable for the complex environment of oil drilling rigs;
[0059] 5. Design an algorithm porting unit to update and adjust the model algorithm based on the vibration and current signals of the motor collected in real time at the drilling site. Compile the motor model and diagnostic algorithm to form a callable, interfaceable, and portable software package, which is then connected to the intelligent operation and maintenance platform for oil drilling equipment to realize the application of motor diagnosis on the drilling platform. Attached Figure Description
[0060] Figure 1 This is a logic block diagram of the fault diagnosis method for variable frequency AC asynchronous motors in oil drilling rigs according to the present invention;
[0061] Figure 2 It is a flowchart of feature extraction and fault identification based on vibration signals;
[0062] Figure 3 (a) and (b) are comparison diagrams of traditional attention mechanism and adaptive sparse module, respectively;
[0063] Figure 4 This is a flowchart of feature extraction and fault identification based on current signals;
[0064] Figure 5 This is a flowchart of the DS evidence theory based on cosine similarity improvement;
[0065] Figure 6 This is a confusion matrix diagram of motor fault diagnosis results;
[0066] Figure 7 This is the logic diagram of the fault diagnosis system for variable frequency AC asynchronous motors in oil drilling rigs according to the present invention. Detailed Implementation
[0067] The present invention will be further described below with reference to the accompanying drawings and embodiments. The drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.
[0068] like Figure 1 As shown, a fault diagnosis method for an AC variable frequency asynchronous motor of an oil drilling rig includes the following steps:
[0069] like Figure 2 The feature vector extraction of the vibration signal includes: inputting the vibration signals of the motor in normal and fault states into a one-dimensional CNN, and extracting the high-dimensional and low-dimensional features of the signal through convolution and pooling operations; in this embodiment, the vibration signals of the motor in the x, y, and z directions are collected for 2 seconds.
[0070] Table 1 shows the parameters of the CNN convolutional network.
[0071] parameter value parameter value Conv1 kernel size 3×3 Conv2 convolution kernel size 3×3 Conv1 convolutional kernel number 8 Conv2 convolutional kernel count 16 Step length 1×1 Batch size 128 Number of learning rounds 2000 Initial learning rate 0.01 Activation function ReLu Optimization function Adam
[0072] First, the vibration signal is segmented, and the segmented input sequence is denoted as x = [x1, x2, ..., x...]. N ], where N is the length of the vibration signal sequence, i.e., the number of segments; the convolution operation for each sequence is defined as follows:
[0073]
[0074] Among them, b and These represent the bias term and the non-linear activation function, respectively; the output z of the convolutional layer i It is by changing the convolution kernel w from x i The process of sliding from the first point to the last point can be viewed as the convolution kernel moving along the corresponding sequence x. i Features learned online.
[0075] Next, the output of the convolutional layer is input into the pooling layer. The pooling operation for each sequence is as follows:
[0076]
[0077]
[0078] Where g is the length of the pool window, p i represents the output of the i-th sequence feature after the pooling operation, and s represents the dimension.
[0079] like Figure 3 As shown, before extracting the feature vectors of vibration signals for fault identification, an adaptive sparse module is designed to dynamically generate attention weights for each feature vector, adaptively calculate the threshold of the attention weights, and filter redundant vectors in real time. The processing method is as follows:
[0080] A threshold estimation module is added to the traditional attention mechanism; the attention layer uses the sigmoid function to scale the scaling parameter to the range of (0,1), thereby obtaining an initial threshold α, as shown in the formula:
[0081] α = sigmoid(W a z+b a ) (twenty one)
[0082] Where z is the output of the attention layer, which is a one-dimensional vector; W a and b a These are the weights and biases, respectively, and 'a' is the corresponding scaling parameter.
[0083] Considering that the threshold not only needs to be positive, but cannot be greater than the maximum absolute value of the attention layer, the formula for the final threshold τ is:
[0084] τ=α·max(z) (22)
[0085] Using the final threshold, a masking function is constructed to filter the output of the attention layer, as shown in the formula:
[0086]
[0087] ReLU is a commonly used activation function that directly sets negative numbers to 0 and retains positive numbers. If a is large enough, b is 0.618. When (z-τ) > 0, the output of f(z) tends to 1. When (z-τ) = 0, the output of f(z) tends to 0 but is greater than 0. When (z-τ) < 0, the output of f(z) is 0.
[0088] Finally, the filtered output and the output of the attention layer are recalculated to reconstruct the output of the attention layer, using the following formula:
[0089]
[0090] Where, N seg Let newz be the sequence number. i This is the output of the sparse attention module.
[0091] Fault identification of vibration signals includes: dividing the filtered vibration signals into 160 training sets and 40 test sets; and analyzing the feature vectors of the training sets into newz... i The input is fed into the Softmax layer. Training is complete when the diagnostic accuracy converges. Then, the feature vectors of the test set are input into the Softmax layer for testing, and the fault identification results are output.
[0092] like Figure 4 As shown, the feature vector extraction and fault identification method of current signal includes: collecting the current signal of the motor for 2 seconds, and dividing the collected three-phase current under normal and fault conditions of the motor into 120 source domain datasets, 40 target domain training sets and 40 target domain test sets.
[0093] Construct a ResNet deep network and initialize the network's weights and parameters;
[0094] Table 2 Parameters of ResNet Deep Network
[0095]
[0096]
[0097] The source and target domain training data are input into the ResNet deep network to extract deep features from the source and target domains.
[0098] By combining the feature transfer model and sharing the network parameters of the source and target domains, the deep features extracted from the source and target domains are simultaneously mapped to RKHS;
[0099] After the fully connected layer of the network, JMMD is introduced to measure the distance between the feature spaces of the two domains. By minimizing the loss function, the difference in feature distribution between the source and target domains is reduced, enabling the model to extract features with high similarity and update the ResNet deep network parameters during backpropagation.
[0100] The target domain test set data is input into the updated ResNet deep network for feature extraction and outputs the fault identification results.
[0101] The JMMD loss function is calculated as follows:
[0102]
[0103] In the formula, H l Let |L| represent the l-th level RKHS, where |L| is the level number of the corresponding set. For the eigenmap of the tensor product in RKHS, z sl With z tl Let P and Q represent the activations of the source and target domains in layer l, respectively; and let E represent the activations of the source and target domains, respectively. p EQ This is the mean distance between the vectors in the source and target domains.
[0104] Embedding JMMD into a feature transfer network yields the following loss function:
[0105] ι=ι c +λι JMMD (D s D t (26)
[0106] In the formula, ι is the loss function of the entire feature transfer network, ι c Let represent the classification loss function of the source domain, and λ be the trade-off parameter for the overall network loss function; D s D t This represents the spatial distance between the source and target domains.
[0107] The fault diagnosis unit uses DS evidence theory to perform decision-level fusion of fault identification results based on vibration and current signals to obtain the final fault diagnosis results, including:
[0108] like Figure 5 Design the DS identification framework, evidence set, and basic probability allocation (BPA) for each evidence source;
[0109] Improving the DS evidence theory using cosine similarity includes:
[0110] Step 1: Calculate the cosine similarity between any two pieces of evidence;
[0111] Among them, the two pieces of evidence are any two of the vibration signals and current signals in the x, y, and z directions;
[0112] Step 2: Cosine similarity between various pieces of evidence (sim) ij Form a similarity matrix S;
[0113]
[0114] Step 3: Calculate the evidence support by summing the similarity scores of all evidence-related items in the matrix.
[0115]
[0116] Define the average similarity of the evidence;
[0117]
[0118] Find the weighting coefficient β i ;
[0119]
[0120] Using weighting coefficient β i To refine the source of evidence; use β i A weighted average of the BPA of all evidence is calculated, and then the weighted average evidence is subjected to n-1 self-combinations using the DS fusion rule to obtain the final fusion result; for example Figure 6 As shown, the overall accuracy rate of fault diagnosis reached 97.5%.
[0121] Table 3 Classification Numbers Corresponding to Motor Status
[0122]
[0123]
[0124] The algorithm porting unit updates and adjusts the model algorithm based on the vibration and current signals of the motor collected in real time at the drilling site, and compiles the motor model and diagnostic algorithm to form a callable, interfaceable, and portable software package, which is seamlessly connected to the intelligent operation and maintenance platform for oil drilling equipment to realize the platform-based application of motor fault diagnosis information.
[0125] like Figure 7 A fault diagnosis system for AC variable frequency asynchronous motors of oil drilling rigs includes:
[0126] The signal acquisition module is used to acquire real-time operating data of the motor at the drilling site, including a vibration sensor to acquire vertical or horizontal vibration signals of the motor; and a current sensor to acquire current signals of the motor.
[0127] The sensor collects vibration and current signals in real time, and after being conditioned by the circuit, it is transmitted to the central processing module via a wireless transmission module.
[0128] The central processing module includes a signal analysis unit, a fault diagnosis unit, and an algorithm porting unit. The signal analysis unit extracts feature vectors from vibration and current signals using CNN and ResNet deep networks, respectively, and performs fault identification to obtain fault identification results based on vibration and current feature vectors. The fault diagnosis unit uses an improved DS evidence theory to perform decision-level fusion of the two identification results to obtain the final fault diagnosis result. The algorithm porting unit ports the motor model, the CNN neural network and ResNet deep network models in the signal analysis unit, and the DS evidence theory model in the fault diagnosis unit to the drilling equipment intelligent operation and maintenance platform, realizing the docking of the algorithm model with the drilling platform, and updating and adjusting the algorithm model according to the real-time acquired vibration and current signals to improve the fault identification accuracy.
[0129] The vibration sensor is installed on the motor housing, and the current sensor is installed on the stator of the motor to obtain the stator current signal.
[0130] The motor failures include static eccentricity, dynamic eccentricity, bearing failure, and rotor bar breakage, which have occurred with the AC variable frequency asynchronous motors of oil drilling rigs.
[0131] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.
Claims
1. A method for diagnosing faults in an AC variable frequency asynchronous motor of an oil drilling rig, characterized in that, Includes the following steps: Step 1: Collect motor vibration signals, input the vibration signals into a CNN neural network, and output fault identification results based on the vibration signals; Step one specifically includes: The collected vibration signals are input into a one-dimensional CNN, and after two convolution and pooling operations, the high-dimensional and low-dimensional features of the signals are extracted. An adaptive sparse module is designed and embedded in a CNN to dynamically generate attention weights for each feature vector output by the CNN front-end network, adaptively calculate the threshold of the attention weights, and filter redundant vectors in real time. The adaptive sparse module includes: A threshold estimation module is added to the attention mechanism. The attention layer uses the sigmoid function to obtain an initial threshold. The formula is: (4) in, z It refers to the output of the attention layer; and These are the weights and biases, respectively. Construct the final threshold The formula is: (5) The masking function is constructed to filter the output of the attention layer, and the formula is as follows: (6) in, relu It is a commonly used activation function; Finally, the filtered output and the output of the attention layer are recalculated to reconstruct the output of the attention layer, resulting in the sparse attention module, as shown in the formula: (7) in, The sequence number; Step 2: Acquire motor current signals, input them into a ResNet deep network, and output fault identification results based on current signals; Step two specifically includes: Step 21: Divide the current signal into source domain samples, target domain training set samples, and target domain test set samples; Step 22: Construct a ResNet deep network and initialize the network's weights and parameters; Step 23: Input the source and target domain training data into the ResNet deep network to extract deep features from the source and target domains; Step 24: Combine the feature transfer model, share the network parameters of the two domains, and simultaneously map the deep features extracted from the source domain and the target domain to RKHS; Step 25: After the fully connected layer of the ResNet deep network, JMMD is introduced to measure the distance between the feature spaces of the two domains. By minimizing the loss function, the difference in feature distribution between the source and target domains is reduced, and the parameters of the ResNet deep network are updated during backpropagation. Step 26: Input the target domain test set data into the updated ResNet deep network for feature extraction and output the fault identification results; Step 3: Input the fault identification results based on current signals and the fault identification results based on vibration signals into the improved DS evidence theory model for decision-level fusion output; Improved DS evidence theory models include: Step 1: Calculate the cosine similarity between any two pieces of evidence; Step 2: Cosine similarity between various pieces of evidence Form a similarity matrix S ; (10) Step 3: Calculate the evidence support by summing the similarity scores of all evidence-related items in the matrix. (11) Define the average similarity of the evidence; (12) Find the weight coefficients ; (13) Using weighting coefficients To correct the source of evidence; using A weighted average of the BPA of all evidence is calculated, and then the DS fusion rule is applied to the weighted average evidence. n -1 self-combination is performed to obtain the final fusion result.
2. The method for diagnosing faults in an AC variable frequency asynchronous motor of an oil drilling rig according to claim 1, characterized in that, Motor fault types include: static eccentricity, dynamic eccentricity, bearing failure, and rotor bar breakage.
3. The method for diagnosing faults in an AC variable frequency asynchronous motor of an oil drilling rig according to claim 1, characterized in that, The JMMD loss function is calculated as follows: (8) In the formula, Indicates the first l Layer RKHS, For the corresponding set level, This represents the eigenmap of the tensor product in RKHS. and The first l Activation of the source and target domains in the layer, , This is the mean distance between the vectors in the source and target domains.
4. The method for diagnosing faults in an AC variable frequency asynchronous motor of an oil drilling rig according to claim 3, characterized in that, Embedding JMMD into the feature transfer network yields the following loss function for the entire feature transfer network: (9) In the formula, The classification loss function represents the source domain. These are the trade-off parameters for the overall network loss function; , This represents the spatial distance between the source and target domains.
5. A system employing the fault diagnosis method for AC variable frequency asynchronous motors of oil drilling rigs according to any one of claims 1-4, characterized in that, include: The system includes a signal acquisition module and a central processing module. The signal acquisition module acquires motor vibration signals and motor current signals through vibration sensors and current sensors, respectively. The central processing module includes a signal analysis unit and a fault diagnosis unit. The signal analysis unit uses CNN neural network and ResNet deep network to extract feature vectors from vibration and current signals, respectively, and performs fault identification. The fault diagnosis unit performs decision-level fusion of the two fault identification results.
6. The fault diagnosis system for AC variable frequency asynchronous motors of oil drilling rigs according to claim 5, characterized in that, The central processing module also includes an algorithm porting unit, which ports the network model from the signal analysis unit to the intelligent operation and maintenance platform for drilling equipment.