Rotating machine state monitoring method and device, readable storage medium and program product

Abstract

The invention provides a rotating machine state monitoring method and device, a readable storage medium and a program product. The method comprises the following steps: acquiring a vibration signal in the operation process of the rotating machine; performing signal conversion and noise reduction processing on the vibration signal to obtain an experimental data set; according to the experimental data set, training to obtain a target rotating machine state monitoring model; and monitoring the state of the rotating machine by using the target rotating machine state monitoring model. According to the scheme, the target rotating machine state monitoring model is determined through the experiment data set, the state of the rotating machine can be monitored in real time, and fault recognition of the rotating machine is facilitated.

Description

Technical Field

[0001] The present application relates to the technical field of condition monitoring, and in particular to a method, device, readable storage medium, and program product for monitoring the condition of rotating machinery. Background Art

[0002] The condition monitoring of rotating machinery is to monitor various status parameters of rotating machinery during operation and understand their changing patterns, so as to determine whether the unit equipment is operating normally, so as to promptly discover possible dangers that may arise during subsequent operation and issue early warnings for possible failures.

[0003] Signal feature extraction technology is a key tool for condition monitoring and fault diagnosis. Mechanical equipment has complex structures and numerous components. The collected dynamic signals represent a comprehensive reflection of the vibrations of each component, and the influence of the transmission path increases the complexity of the signal. During the monitoring process, the status signals collected during system operation are first analyzed and feature data extracted. Then, fault-related symptoms are identified and used for condition monitoring and fault identification.

[0004] The current research still has the following deficiencies:

[0005] 1. Current research in the field of rotating machinery condition monitoring does not meet the accuracy requirements, making it impossible to accurately implement condition monitoring.

[0006] Second, existing technologies cannot achieve the organic coordination and unification of the three key implementation steps of signal analysis, feature extraction, and state monitoring, resulting in poor system matching.

[0007] 3. Existing technologies focus on theoretical research and ignore experimental conditions under actual working conditions, resulting in the proposed methods having good effects in cases but poor effects under actual conditions. Summary of the Invention

[0008] The purpose of this application is to provide a rotating machinery state monitoring method, device, readable storage medium and program product to solve the drawbacks of rotating machinery state monitoring in the prior art.

[0009] In a first aspect, an embodiment of the present application provides a method for monitoring the state of a rotating machine, comprising:

[0010] Obtain vibration signals during the operation of rotating machinery;

[0011] Performing signal conversion and noise reduction processing on the vibration signal to obtain an experimental data set;

[0012] According to the experimental data set, a target rotating machinery state monitoring model is trained;

[0013] The target rotating machinery state monitoring model is used to monitor the state of the rotating machinery.

[0014] Optionally, training a target rotating machinery state monitoring model based on the experimental data set includes:

[0015] Dividing the experimental data set into a training set and a validation set according to a preset ratio;

[0016] Inputting the training set into the primary feature extraction module of the initial state monitoring model to obtain first feature data; the first feature data is used to represent low-dimensional change trend feature information of the data;

[0017] Inputting the first feature data into the deep feature extraction module of the initial state monitoring model to obtain second feature data; the second feature data is used to represent the correlation characteristics between the previous and next time steps of the data in high dimension;

[0018] Inputting the second characteristic data into the mapping relationship building module of the initial state monitoring model to determine and build a nonlinear mapping relationship between the vibration signal of the rotating machinery and the health state;

[0019] Determining a trained first monitoring model according to the nonlinear mapping relationship;

[0020] Inputting the verification set into the first monitoring model to determine whether an output result of the first monitoring model meets a convergence condition;

[0021] If the output result does not meet the convergence condition, the first monitoring model is iterated using the training set until the output result meets the convergence condition, and the current first monitoring model is determined to be the target rotating machinery state monitoring model.

[0022] Optionally, the primary feature extraction module includes a one-dimensional convolution layer, a batch normalization layer, a first activation function layer, and an average pooling layer. Inputting the training set into the primary feature extraction module of the initial state monitoring model to obtain first feature data includes:

[0023] Inputting the training set into the one-dimensional convolutional layer to perform primary feature extraction and a first dimensionality reduction operation to determine a first eigenvector;

[0024] The first feature vector is sequentially input into the first activation function layer, the first activation function layer and the average pooling layer to obtain the first feature data after the second dimensionality reduction operation.

[0025] Optionally, the deep feature extraction module includes at least three attention mechanism layers, a second activation function layer, and a maximum pooling layer. Inputting the first feature data into the deep feature extraction module to obtain the second feature data includes:

[0026] Inputting the first feature data into the at least three attention mechanism layers in sequence to determine the processed hidden feature information and key feature information;

[0027] The hidden feature information and the key feature information are sequentially input into the second activation function layer and the maximum pooling layer to obtain second feature data.

[0028] Optionally, inputting the first feature data into the at least three attention mechanism layers in sequence to determine the processed hidden feature information and key feature information includes:

[0029] The first feature data is sequentially input into the at least three attention mechanism layers, and the processed hidden feature information and key feature information are determined by a first formula, wherein the first formula is expressed as:

[0030]

[0031] Where X represents a vector sequence of N elements in the first feature data; q represents a query vector; α n represents the weight of the nth element, and

[0032] Optionally, inputting the second characteristic data into a mapping relationship building module of the initial state monitoring model to determine and build a nonlinear mapping relationship between the vibration signal of the rotating machinery and the health state includes:

[0033] Inputting the second characteristic data into the temporal convolutional neural network of the mapping relationship building module to determine and build a nonlinear mapping relationship between the vibration signal of the rotating machinery and the health status;

[0034] The temporal convolutional neural network comprises at least: an expanded causal convolution layer, a weight normalization layer, a clipping function layer, a third activation function layer and a random inactivation dropout layer.

[0035] Optionally, performing signal conversion and noise reduction processing on the vibration signal to obtain an experimental data set includes:

[0036] Performing signal conversion on the vibration signal to determine a digital signal of the vibration signal;

[0037] performing time domain characteristic analysis, data defect preprocessing and filling, and data noise reduction processing on the digital signal to determine processed first data; the time domain characteristic analysis includes calculating at least one time domain characteristic of the digital signal including an average value, variance, standard deviation, and peak value, performing feature extraction with time as a variable, and determining a characteristic of the digital signal changing over time;

[0038] The first data is formatted in a unified manner to determine a plurality of groups of processed data samples, wherein the plurality of groups of data samples have a unified dimension and length;

[0039] Each group of data samples is assigned a data label to obtain an experimental data set.

[0040] In a second aspect, the present application provides a rotating machinery condition monitoring device, comprising:

[0041] An acquisition module is used to acquire vibration signals during the operation of the rotating machinery;

[0042] A first determination module is configured to perform signal conversion and noise reduction processing on the vibration signal to obtain an experimental data set;

[0043] A second determination module is configured to train a target rotating machinery state monitoring model based on the experimental data set;

[0044] The first processing module is configured to monitor the state of the rotating machinery using the target rotating machinery state monitoring model.

[0045] In a third aspect, the present application provides a readable storage medium having a program or instruction stored thereon, which, when executed by a processor, implements the steps in the method described in the first aspect.

[0046] In a fourth aspect, the present application provides a computer program product, comprising computer instructions, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0047] The beneficial effects of the above technical solution of this application are as follows:

[0048] The rotating machinery state monitoring method, device, readable storage medium and program product of the embodiments of the present application obtain an experimental data set by performing signal conversion and noise reduction processing on the vibration signal obtained during the operation of the rotating machinery, and use the experimental data set to train a target rotating machinery state monitoring model. The target rotating machinery state monitoring model is used to perform real-time rotating machinery state monitoring, which can improve the accuracy of rotating machinery state monitoring and effectively achieve the organic coordination and unity of signal analysis, feature extraction and state monitoring. In addition, the effect of the target rotating machinery state monitoring model generated based on the vibration signal under actual working conditions is significantly improved compared with the experimental data of theoretical research. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A schematic diagram of a flow chart of a rotating machinery state monitoring method provided in an embodiment of the present application;

[0050] Figure 2 A schematic diagram of the structure of the initial state monitoring model provided in an embodiment of the present application;

[0051] Figure 3 An application flow chart of the rotating machinery state monitoring method provided in an embodiment of the present application;

[0052] Figure 4 This is a schematic diagram of the structure of a rotating machinery state monitoring device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0053] In order to make the technical problems, technical solutions and advantages to be solved by this application clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0054] It should be understood that references throughout this specification to "one embodiment" or "an embodiment" mean that a particular feature, structure, or characteristic associated with the embodiment is included in at least one embodiment of the present application. Therefore, the appearances of "in one embodiment" or "in an embodiment" throughout this specification do not necessarily refer to the same embodiment. Furthermore, these particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0055] In the embodiments of this application, the term "and / or" describes the association relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally indicates that the associated objects are in an "or" relationship.

[0056] In the embodiments of the present application, the term "plurality" refers to two or more than two, and other quantifiers are similar.

[0057] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0058] The present invention provides a method, apparatus, readable storage medium, and program product for monitoring the condition of rotating machinery. The method and apparatus are based on the same patent application concept. Since the method and apparatus solve similar problems, the implementation of the apparatus and method can refer to each other, and any repetitions will not be repeated.

[0059] like Figure 1 As shown, the rotating machinery state monitoring method provided by the embodiment of the present application includes:

[0060] Step 11: Acquire vibration signals during the operation of the rotating machinery.

[0061] This application uses a three-axis acceleration sensor to collect vibration signals during the actual operation of a rotating machine. For example, a three-axis acceleration sensor is used to collect vibration signals during the operation of a rotating machine. The acceleration sensor collects data in the X, Y, and Z directions as acceleration signals.

[0062] Step 12: Perform signal conversion and noise reduction on the vibration signal to obtain an experimental data set.

[0063] This application performs signal conversion on the vibration signal to standardize and normalize the vibration signal, and uses the continuous wavelet transform method to reduce the noise of the data set after standardization and normalization to obtain an experimental data set. The experimental data set provides time scale information of the signal, can extract features that change over time, and is suitable for analyzing transient or non-stationary signals.

[0064] Step 13: training a target rotating machinery state monitoring model based on the experimental data set;

[0065] Step 14: Monitor the state of the rotating machinery using the target rotating machinery state monitoring model.

[0066] In an embodiment of the present application, a target rotating machinery state monitoring model is generated by training using an experimental data set. This model can be used to monitor the state of rotating machinery, which is conducive to obtaining accurate vibration signals and effectively realizing the organic coordination and unity of signal analysis, feature extraction, and state monitoring. Moreover, the effect of the target rotating machinery state monitoring model generated based on the vibration signals under actual working conditions is significantly improved compared with the experimental data of theoretical research.

[0067] Optionally, the rotating machinery includes but is not limited to motors, generators, gearboxes, pumps, and fans. The vibration signal is but is not limited to acceleration signals, speed signals, displacement signals, torque signals, and sound pressure signals of the rotating machinery.

[0068] Optionally, in the embodiment of the present application, the above step 12 includes:

[0069] Performing signal conversion on the vibration signal to determine a digital signal of the vibration signal;

[0070] performing time domain characteristic analysis, data defect preprocessing and filling, and data noise reduction processing on the digital signal to determine processed first data; the time domain characteristic analysis includes calculating at least one time domain characteristic of the digital signal including an average value, variance, standard deviation, and peak value, performing feature extraction with time as a variable, and determining a characteristic of the digital signal changing over time;

[0071] The first data is formatted in a unified manner to determine a plurality of groups of processed data samples, wherein the plurality of groups of data samples have a unified dimension and length;

[0072] Each group of data samples is assigned a data label to obtain an experimental data set.

[0073] In an embodiment of the present application, the vibration signal is converted to obtain a visual digital signal. For example, an edge computing system can be built, a data acquisition unit is arranged, and a multi-channel signal acquisition card, an analog / digital (A / D) conversion module, and a signal conditioning module are integrated. The high-speed acquisition card processes the vibration signal collected by the acceleration sensor, and the sampling rate is set to 2kHz. The acceleration signal sensor uses an IEPE (integrated circuit voltage sensor) signal conditioning module to further reduce line noise information by external power supply.

[0074] Furthermore, the digital signal is subjected to noise reduction processing. First, the application performs time domain characteristic analysis, visually inspects the digital signal data set in chronological order, and calculates the time domain characteristics of the digital signal data set, such as the mean, variance, standard deviation, and peak value. The data is extracted using time as a variable to find the trend of the signal curve, characteristic information, and the changing trend of information in the time dimension, and determine the changing characteristics of the digital signal over time.

[0075] The data after time domain characteristic analysis is then preprocessed to fill in data defects. This application uses the random forest algorithm to fill in missing values ​​in the data set. The specific operation process is as follows:

[0076] (1) Input: Set the training set S = {(x i ,y i ),i=1,2,…,n},(X,Y)∈R d ×R: test sample X i ∈R d .

[0077] (2) Output: Tree T = {h i (x i ),i=1,2,…,N tree}.

[0078] (3) Cycle through the tree T until a random forest is generated. The cyclic processing flow is as follows: Use the replacement resampling method to obtain the training set S from the original training set. i , randomly select M from d feature attribute values try characteristic attribute values, according to the Gini index from each node M try Select the best feature value among the feature attribute values, and use S i Generate the corresponding unpruned number h i , split until the tree grows to the maximum, generating a random forest, and ending.

[0079] After pre-processing and filling data defects, this application performs data noise reduction. This application can use the continuous wavelet transform method to reduce the noise of the data set. This method provides the time scale information of the signal and can extract the characteristics that change over time. It is suitable for analyzing transient or non-stationary signal data sets. The continuous wavelet transform formula is shown in the following formula (1):

[0080]

[0081] In formula (1), Represents the mother wavelet; parameter a represents the scale coefficient, which is the inverse of the frequency; parameter b represents the time shift.

[0082] The discrete wavelet transform is derived from the discretization of the continuous wavelet transform. This application uses binary discretization, as shown in the following formula (2):

[0083]

[0084] In formula (2), a and b are replaced by 2j and 2jk. Discrete wavelet transform can accurately characterize signal characteristics through wavelet basis functions transformed at different scales, and continuously focus on any small details of the signal.

[0085] After performing time domain characteristic analysis, data defect preprocessing and filling, and data noise reduction, the present application obtains processed first data, organizes the first data into a unified format, and determines multiple groups of processed data samples; assigns a data label to each group of data samples to obtain an experimental data set, that is, organizes each experimental data sample, assigns a label, and establishes an experimental data set. The label indicates a first sample label that indicates the rotating mechanical equipment is in a normal operating state, and a second sample label that indicates the rotating mechanical equipment is in an abnormal state.

[0086] In particular, rotating machinery can be turbines, generators, pumps, and other equipment. The sample set consists of historical fault data for these equipment, including 15 operating states: one normal state and 14 common fault states. In this case, the sample set has 16 sample categories, each of which is denoted by a unique label. These sample labels are used to characterize the equipment's fault type and are used for subsequent data processing and model training. Furthermore, each sample in this sample set for each operating state includes 48 characteristic parameters, including parameters such as the equipment's output voltage, output current, and input current.

[0087] For example, this application will organize the first data after time domain characteristic analysis, random forest missing value filling and data noise reduction processing, mainly to unify the signal format in the first data. Since the signal data length in each group of data samples is different, it cannot be smoothly input into the subsequent initial state monitoring model. Therefore, it is necessary to unify the dimension and length of each group of data samples, organize and divide them uniformly, and match the processed data with the labels one by one to establish an experimental data set.

[0088] Furthermore, step 13 of the present application includes:

[0089] Dividing the experimental data set into a training set and a validation set according to a preset ratio;

[0090] Inputting the training set into the primary feature extraction module of the initial state monitoring model to obtain first feature data; the first feature data is used to represent low-dimensional change trend feature information of the data;

[0091] Inputting the first feature data into the deep feature extraction module of the initial state monitoring model to obtain second feature data; the second feature data is used to represent the correlation characteristics between the previous and next time steps of the data in high dimension;

[0092] Inputting the second characteristic data into the mapping relationship building module of the initial state monitoring model to determine and build a nonlinear mapping relationship between the vibration signal of the rotating machinery and the health state;

[0093] Determining a trained first monitoring model according to the nonlinear mapping relationship;

[0094] Inputting the verification set into the first monitoring model to determine whether an output result of the first monitoring model meets a convergence condition;

[0095] If the output result does not meet the convergence condition, the first monitoring model is iterated using the training set until the output result meets the convergence condition, and the current first monitoring model is determined to be the target rotating machinery state monitoring model.

[0096] In an embodiment of the present application, the preset ratio can be 7.5:2.5, that is, the present application divides the experimental data set into a training set and a verification set according to the ratio of 7.5:2.5, and uses the training set to input into the initial state monitoring model to obtain the trained first monitoring model, and uses the verification set to verify whether the output result of the first monitoring model meets the convergence condition. The convergence condition is to determine whether the output result is within the preset range. If the convergence condition is met, it is determined that the current first monitoring model can be used for rotating machinery state monitoring. Otherwise, the first monitoring model is retrained using the training set until the latest model meets the convergence condition, and then the current model is saved and output.

[0097] The initial state monitoring model of the present application includes: a primary feature extraction module, a deep feature extraction module, and a mapping relationship building module. The primary feature extraction module is used to perform primary feature extraction to obtain first feature data for representing the low-dimensional change trend feature information of the data; the first feature data and the deep feature extraction module integrated with the attention mechanism are used to perform deep feature extraction to obtain second feature data for representing the correlation characteristics between the previous and next time steps of the high-dimensional data; the second feature data is input into the mapping relationship building module of the initial state monitoring model to determine the nonlinear mapping relationship between the vibration signal of the rotating machinery and the health status.

[0098] Specifically, refer to Figure 2 The schematic diagram of the initial state monitoring model shown in the figure shows that after the experimental dataset is sequentially input into the primary feature extraction module, the deep feature extraction module, and the mapping relationship construction module, a state monitoring value can be obtained, which is used to characterize the state of the rotating machinery. The primary feature extraction module includes a one-dimensional convolution (1DCNN) layer, a batch normalization (BatchNorm) layer, a first activation function (ReLu) layer, and an average pooling layer; the deep feature extraction module includes at least three attention mechanism layers (Attention layer), a second activation function (ReLu) layer, and a maximum pooling layer (Max pooling); and the mapping relationship construction module includes a dilated causal convolution layer (Dilated CausalConv), a weight normalization (WeightNorm) layer, a chomping function (Chomp) layer, a third activation function (ReLu) layer, and a random dropout (dropout) layer.

[0099] Optionally, the primary feature extraction module includes a one-dimensional convolution layer, a batch normalization layer, a first activation function layer, and an average pooling layer. Inputting the training set into the primary feature extraction module of the initial state monitoring model to obtain first feature data includes:

[0100] Inputting the training set into the one-dimensional convolutional layer to perform primary feature extraction and a first dimensionality reduction operation to determine a first eigenvector;

[0101] The first feature vector is sequentially input into the first activation function layer, the first activation function layer and the average pooling layer to obtain the first feature data after the second dimensionality reduction operation.

[0102] In the embodiment of the present application, the primary feature extraction module consists of a one-dimensional convolution layer, a batch normalization layer, an activation function layer, and an average pooling layer. All the information input each time first undergoes the primary feature extraction and dimensionality reduction operation of the one-dimensional convolution layer. The calculation process is shown in formula (3):

[0103]

[0104] Wherein, in formula (3): M j Represents the input data set; l represents the lth layer of the network; x represents the activation function value of the feature vector; k represents the convolution kernel vector of the feature vector; * represents the convolution operation; b represents the deviation of the feature vector; f represents the activation function.

[0105] After the primary feature extraction and the first dimensionality reduction operation in the one-dimensional convolution layer, the first eigenvector is determined and input into the batch normalization layer to realize the standardization of the input data, reduce the difference between different batches, and speed up the convergence. The calculation process is shown in formula (4):

[0106]

[0107] In formula (4), represents the training batch; E(x (k) ) represents the mean of the training set; var(x (k) ) represents the training set variance.

[0108] After being processed by the first activation function layer, the fitting ability of the neural network is enhanced, the effective search range of the parameter space is effectively suppressed, the calculation amount of the model is reduced, and the calculation efficiency of back propagation is effectively increased;

[0109] Finally, after the processing of the average pooling layer, the extracted primary features are subjected to dimensionality reduction again, that is, a second dimensionality reduction operation is performed to reduce the number of network parameters. The primary feature extraction module of this application uses average pooling calculation, and the pooling calculation process is shown in formula (5):

[0110]

[0111] In formula (5), down(·) represents the downsampling function; β is the connection weight; b represents the bias of the layer; and f(·) represents the activation function of the pooling layer.

[0112] Optionally, the deep feature extraction module includes at least three attention mechanism layers, a second activation function layer, and a maximum pooling layer. Inputting the first feature data into the deep feature extraction module to obtain the second feature data includes:

[0113] Inputting the first feature data into the at least three attention mechanism layers in sequence to determine the processed hidden feature information and key feature information;

[0114] The hidden feature information and the key feature information are sequentially input into the second activation function layer and the maximum pooling layer to obtain second feature data.

[0115] In an embodiment of the present application, the first feature data is sequentially input into the at least three layers of attention mechanism layers to determine the hidden feature information and key feature information after processing, that is, the first feature data is input into the deep feature extraction module for processing. The deep feature extraction module performs deep mining and data dimensionality reduction on all input feature data, further refines the feature data that can characterize the real degradation information, and performs feature dimensionality reduction on this basis. The deep feature extraction module consists of a three-layer attention mechanism, an activation function, and a maximum pooling layer. The feature data is calculated by the three-layer attention mechanism in sequence, which can fully mine the hidden feature information and key feature information in the data, and then processed by the activation function and the maximum pooling layer to obtain the second feature data for representing the correlation characteristics between the previous and next time steps of the high-dimensional data.

[0116] Specifically, the first feature data is sequentially input into the at least three attention mechanism layers to determine the processed hidden feature information and key feature information, including:

[0117] The first feature data is sequentially input into the at least three attention mechanism layers, and the processed hidden feature information and key feature information are determined by a first formula, wherein the first formula is expressed as:

[0118]

[0119] Where X represents a vector sequence of N elements in the first feature data; q represents a query vector; α n represents the weight of the nth element, and

[0120] X in this application can also be expressed as X=[x1, x2, ..., x N ]∈R d×N is a vector sequence containing N elements, q∈R d is a query vector, α n is the weight of the nth element, satisfying

[0121] In the network model adopted in this application, the three-layer attention mechanism can adaptively improve the network's sensitivity to different channels, thereby optimizing network performance; at the same time, in actual working scenarios, rotating mechanical movement often contains random interference. The three-layer attention mechanism can enable the network model to automatically focus on the vibration signal containing fault characteristics, effectively extract the fault characteristics, and greatly improve the diagnostic accuracy and training speed.

[0122] Optionally, this application can add an attention threshold structure to each layer of the attention mechanism layer, and use this attention threshold structure to determine whether performance degradation has occurred; select the activation function, loss function and optimization algorithm suitable for the model, so that before inputting the time convolutional neural network (mapping relationship building module), it can better solve nonlinear problems with complex data characteristics and minimize the loss of the time convolutional neural network. The specific implementation steps of the attention threshold function of this application are as follows:

[0123] 1) Data input: input the performance status value at that moment;

[0124] 2) Calculate the difference: Calculate the difference between the performance status value at this moment and the performance status value at the previous moment;

[0125] 3) Determine whether it is a performance degradation point: Determine the magnitude relationship between the value and the degradation threshold at that time. At that time, it is determined that the rotating machinery has experienced performance degradation at that time. At that time, it is determined that the rotating machinery has not experienced performance degradation at that time.

[0126] 4. Change coefficient: When the moment is a performance degradation point, the influence of the moment state value on model training is increased by multiplying the moment performance characteristic value by the influence factor. The initial value of the influence factor is 1, and the value is adjusted later through model training. If the moment does not experience performance degradation, it remains unchanged.

[0127] 5) Input the performance status value into the temporal convolutional neural network for training.

[0128] Optionally, inputting the second characteristic data into a mapping relationship building module of the initial state monitoring model to determine and build a nonlinear mapping relationship between the vibration signal of the rotating machinery and the health state includes:

[0129] Inputting the second characteristic data into the temporal convolutional neural network of the mapping relationship building module to determine and build a nonlinear mapping relationship between the vibration signal of the rotating machinery and the health status;

[0130] The temporal convolutional neural network comprises at least: an expanded causal convolution layer, a weight normalization layer, a clipping function layer, a third activation function layer and a random inactivation dropout layer.

[0131] It should be noted that the dropout layer is achieved by traversing the nodes of each layer of the neural network, and then setting a keep_prob (node ​​retention probability) for the neural network of this layer, that is, the nodes of this layer have a keep_prob probability of being retained, and the value range of keep_prob is between 0 and 1. By setting the keep_prob probability of the nodes of this layer of the neural network, the neural network will not be biased towards a certain node (because the node may be deleted), so that the weight of each node will not be too large, thereby reducing the overfitting of the neural network.

[0132] In the embodiment of the present application, the second feature data is input into the time convolutional neural network of the mapping relationship building module. The time convolutional neural network includes at least: an expansion causal convolution layer, a weight normalization layer, a clipping function layer, a third activation function layer and a random inactivation dropout layer. The specific structure can be adopted Figure 2 The structure shown in FIG1 is used to determine the nonlinear mapping relationship between the vibration signal of the rotating machinery and the health status through a time convolutional neural network. The main function of the mapping relationship building module of the present application is to establish a nonlinear mapping relationship between the front and back data based on the multi-level features extracted previously. The mapping relationship building module of the present application is mainly calculated by the dilated causal convolutional network. The calculation process of the dilated convolution is shown in formula (6):

[0133]

[0134] Wherein, in formula (6): f(γ) is the element at position γ in the convolution kernel; x t-d·i is the sequence data sampled by dilated convolution; t is the output of the dilated convolution at time t; d is the dilation coefficient; k is the size of the convolution kernel.

[0135] Optionally, the present application iteratively trains the initial state monitoring model. When the model's monitoring error is within a specified range, the model parameters are saved and used for rotating machinery state monitoring in the current operating condition. The current model parameters are fine-tuned using transfer learning methods in rotating machinery state monitoring under similar operating conditions to achieve cross-operating condition monitoring.

[0136] Optionally, the target rotating machinery status monitoring model of the present application can be connected to a cloud system. By monitoring the status of the rotating machinery, the data can be uploaded to the cloud in real time. When the current status is judged to be unhealthy, the fault can be marked on the cloud, such as triggering an event or log. After the event is captured by the external system, a maintenance notification is sent to the equipment maintenance personnel via SMS, application notification or instant messaging tool. The maintenance notification includes the fault equipment ID, fault information, timestamp and preliminary fault analysis.

[0137] For example, after an event is captured by an external system, the automation system sends a notification to the relevant equipment maintenance personnel via SMS, application notification, or other instant messaging tools. The notification contains the faulty equipment ID, detailed fault information, timestamp, and preliminary fault analysis; the maintenance personnel analyze the fault data on the blockchain and combine it with other historical data and sensor data readings to determine the cause of the fault and the best repair strategy; after the fault is repaired, a new transaction is sent to the blockchain, recording the details of the repair, the spare parts used, the maintenance personnel, and the completion time.

[0138] Reference Figure 3 As shown, the embodiment of the present application provides a specific implementation method, which adopts a temporal convolutional network with an attention mechanism to monitor the state of rotating machinery, which includes the following steps:

[0139] (1) Collect vibration signal data of the rotating machinery during operation through a three-axis acceleration sensor;

[0140] (2) Data format conversion and processing. The vibration signal is converted into data format and processed through the high-speed acquisition card, A / D conversion module, and signal conditioning module in sequence;

[0141] (3) Extract the time domain features of the signal data and extract the dimensionless and dimensionless parameters: cutting force mean, variance, standard deviation, root mean square value, peak-to-peak value and skewness as experimental data;

[0142] (4) Construct a rotating machinery state monitoring model with a temporal convolutional network that incorporates an attention mechanism;

[0143] (5) Use experimental data to continuously iteratively train the above monitoring model. When the model state monitoring error is below the specified threshold, save the current model parameters as experimental model parameters;

[0144] (6) Based on the experimental model parameters, a temporal convolutional network with attention mechanism is used to monitor the state of rotating machinery in real time and output the real-time health status.

[0145] The method of this application can extract the time domain feature parameters of the corresponding signal data segment based on actual three-dimensional vibration signal data, perform corresponding feature enhancement at the data dimension level, and strengthen the differential characteristics between different health states; based on the adaptive feature extraction, this application incorporates the attention mechanism for deep feature extraction, further enhancing the differentiation of time domain features under different health states, and fully mining the characteristics of the vibration signal. This application selects the network layer and corresponding settings of the corresponding parameter magnitude based on the data magnitude and scale, so that the correspondence between the network and the data set to be learned is stronger, which can avoid the problems of gradient disappearance and network degradation.

[0146] Compared with the prior art, the present invention has the following advantages:

[0147] The rotating machinery state monitoring method involved in the present application has high accuracy and can accurately identify the current health status type of the rotating machinery; the rotating machinery state monitoring model proposed by the method of the present application starts from the source of the fault diagnosis task at each stage involved in the entire life cycle of the rotating machinery, focuses on the two steps of signal analysis and feature extraction, and improves the robustness and generalization of the model; the present application provides technical support for the intelligent development of rotating machinery state monitoring. Through the rotating machinery state monitoring method, users can monitor the current real-time state of the rotating machinery in real time, providing convenient conditions for the health status management of the equipment.

[0148] Reference Figure 4 As shown, the embodiment of the present application further provides a rotating machinery state monitoring device, comprising:

[0149] An acquisition module 41 is used to acquire a vibration signal during the operation of the rotating machinery;

[0150] A first determination module 42 is configured to perform signal conversion and noise reduction processing on the vibration signal to obtain an experimental data set;

[0151] A second determining module 43 is configured to train a target rotating machinery state monitoring model based on the experimental data set;

[0152] The first processing module 44 is configured to monitor the state of the rotating machinery using the target rotating machinery state monitoring model.

[0153] Optionally, in the embodiment of the present application, the second determining module 43 includes:

[0154] A first processing unit, configured to divide the experimental data set into a training set and a validation set according to a preset ratio;

[0155] A first acquisition unit is configured to input the training set into a primary feature extraction module of an initial state monitoring model to acquire first feature data; the first feature data is used to represent low-dimensional change trend feature information of the data;

[0156] A second acquisition unit is configured to input the first feature data into a deep feature extraction module of the initial state monitoring model to acquire second feature data; the second feature data is used to represent the correlation characteristics between the previous and next time steps of the data in high dimension;

[0157] a first determining unit, configured to input the second characteristic data into a mapping relationship building module of the initial state monitoring model, and determine to build a nonlinear mapping relationship between the vibration signal of the rotating machinery and the health state;

[0158] A second determining unit, configured to determine a trained first monitoring model according to the nonlinear mapping relationship;

[0159] a third determining unit, configured to input the verification set into the first monitoring model and determine whether an output result of the first monitoring model satisfies a convergence condition;

[0160] The second processing unit is configured to iterate the first monitoring model using the training set if the output result does not meet the convergence condition, and determine that the current first monitoring model is the target rotating machinery state monitoring model when the output result meets the convergence condition.

[0161] Optionally, the primary feature extraction module includes a one-dimensional convolution layer, a batch normalization layer, a first activation function layer and an average pooling layer, and the first acquisition unit includes:

[0162] A first determining subunit is configured to input the training set into the one-dimensional convolutional layer to perform primary feature extraction and a first dimensionality reduction operation, and determine a first eigenvector;

[0163] The second determining subunit is used to input the first feature vector into the first activation function layer, the first activation function layer and the average pooling layer in sequence to obtain the first feature data after the second dimensionality reduction operation.

[0164] Optionally, the deep feature extraction module includes at least three attention mechanism layers, a second activation function layer and a maximum pooling layer, and the second acquisition unit includes:

[0165] A third determining subunit is configured to sequentially input the first feature data into the at least three attention mechanism layers to determine the processed hidden feature information and key feature information;

[0166] The first acquisition subunit is used to input the hidden feature information and the key feature information into the second activation function layer and the maximum pooling layer in sequence to obtain second feature data.

[0167] Optionally, in the embodiment of the present application, the third determining subunit is specifically configured to:

[0168] The first feature data is sequentially input into the at least three attention mechanism layers, and the processed hidden feature information and key feature information are determined by a first formula, wherein the first formula is expressed as:

[0169]

[0170] Where X represents a vector sequence of N elements in the first feature data; q represents a query vector; αn represents the weight of the nth element, and

[0171] Optionally, in the embodiment of the present application, the first determining unit includes:

[0172] a fourth determining subunit, configured to input the second feature data into the temporal convolutional neural network of the mapping relationship building module to determine and build a nonlinear mapping relationship between the vibration signal of the rotating machinery and the health status;

[0173] The temporal convolutional neural network comprises at least: an expanded causal convolution layer, a weight normalization layer, a clipping function layer, a third activation function layer and a random inactivation dropout layer.

[0174] Optionally, in the embodiment of the present application, the first determining module 42 includes:

[0175] a fourth determining unit, configured to perform signal conversion on the vibration signal to determine a digital signal of the vibration signal;

[0176] a fifth determining unit, configured to perform time domain characteristic analysis, data defect preprocessing and filling, and data noise reduction on the digital signal to determine processed first data; the time domain characteristic analysis includes calculating at least one time domain characteristic of the digital signal: mean, variance, standard deviation, and peak value; performing feature extraction with time as a variable, and determining a characteristic of the digital signal changing over time;

[0177] a sixth determining unit, configured to format the first data in a unified format and determine a plurality of groups of data samples after processing; wherein the plurality of groups of data samples have a unified dimension and length;

[0178] The third acquisition unit is used to assign a data label to each group of data samples to obtain an experimental data set.

[0179] Among them, the implementation embodiments of the above-mentioned rotating machinery state monitoring method are all applicable to the embodiments of the rotating machinery state monitoring device and can also achieve the same technical effects.

[0180] A readable storage medium in an embodiment of the present application stores a program or instruction thereon. When the program or instruction is executed by a processor, the steps in the rotating machinery state monitoring method as described above are implemented and the same technical effect can be achieved. To avoid repetition, they will not be described here.

[0181] The processor is the processor in the rotating machinery condition monitoring method described in the above embodiment. The readable storage medium includes a computer-readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0182] The present application also provides a computer program product including computer instructions, which, when executed by a processor, implement the above Figure 1 The various processes of the method embodiment shown can achieve the same technical effect, and to avoid repetition, they will not be described here.

[0183] In embodiments of the present application, modules can be implemented in software so that they can be executed by various types of processors. For example, an identified executable code module can include one or more physical or logical blocks of computer instructions, for example, which can be constructed as objects, processes, or functions. Nevertheless, the executable code of the identified module does not need to be physically located together, but can include different instructions stored in different locations, which, when logically combined together, constitute the module and achieve the specified purpose of the module.

[0184] In fact, executable code module can be a single instruction or many instructions, and can even be distributed on a plurality of different code segments, distributed in the middle of different programs, and distributed across a plurality of memory devices.Similarly, operating data can be identified in the module, and can be implemented and organized in the data structure of any appropriate type according to any appropriate form.Described operating data can be collected as a single data set, or can be distributed in different locations (including on different storage devices), and can only be present on a system or network as an electronic signal at least in part.

[0185] When a module can be implemented using software, given the current state of hardware technology, those skilled in the art can build corresponding hardware circuits to implement the corresponding functions of the module, regardless of cost. The hardware circuits may include conventional very large scale integration (VLSI) circuits or gate arrays, as well as existing semiconductors such as logic chips and transistors, or other discrete components. Modules may also be implemented using programmable hardware devices, such as field programmable gate arrays, programmable array logic, or programmable logic devices.

[0186] The above exemplary embodiments are described with reference to the accompanying drawings. Many different forms and embodiments are possible without departing from the spirit and teachings of this application. Therefore, this application should not be construed as limited to the exemplary embodiments set forth herein. Rather, these exemplary embodiments are provided so that this application will be complete and impartial and will convey the scope of this application to those skilled in the art. In the drawings, component sizes and relative sizes may be exaggerated for clarity. The terminology used herein is for purposes of describing specific exemplary embodiments only and is not intended to be limiting. As used herein, unless the context clearly indicates otherwise, the singular forms "a," "an," and "the" are intended to encompass such plural forms. It will be further understood that the terms "comprising" and / or "including," when used in this specification, indicate the presence of the stated features, integers, steps, operations, components, and / or elements, but do not preclude the presence or addition of one or more other features, integers, steps, operations, components, elements, and / or groups thereof. Unless otherwise indicated, when stated, a range of values ​​includes the upper and lower limits of that range and any subranges therebetween.

[0187] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for monitoring the state of a rotating machine, characterized in that: include: Obtain vibration signals during the operation of rotating machinery; Performing signal conversion and noise reduction processing on the vibration signal to obtain an experimental data set; According to the experimental data set, a target rotating machinery state monitoring model is trained; The target rotating machinery state monitoring model is used to monitor the state of the rotating machinery.

2. The method according to claim 1, characterized in that Based on the experimental data set, a target rotating machinery condition monitoring model is trained, including: Dividing the experimental data set into a training set and a validation set according to a preset ratio; Inputting the training set into the primary feature extraction module of the initial state monitoring model to obtain first feature data; the first feature data is used to represent low-dimensional change trend feature information of the data; Inputting the first feature data into the deep feature extraction module of the initial state monitoring model to obtain second feature data; the second feature data is used to represent the correlation characteristics between the previous and next time steps of the data in high dimension; Inputting the second characteristic data into the mapping relationship building module of the initial state monitoring model to determine and build a nonlinear mapping relationship between the vibration signal of the rotating machinery and the health state; Determining a trained first monitoring model according to the nonlinear mapping relationship; Inputting the verification set into the first monitoring model to determine whether an output result of the first monitoring model meets a convergence condition; If the output result does not meet the convergence condition, the first monitoring model is iterated using the training set until the output result meets the convergence condition, and the current first monitoring model is determined to be the target rotating machinery state monitoring model.

3. The method according to claim 2, characterized in that The primary feature extraction module includes a one-dimensional convolution layer, a batch normalization layer, a first activation function layer, and an average pooling layer. The training set is input into the primary feature extraction module of the initial state monitoring model to obtain first feature data, including: Inputting the training set into the one-dimensional convolutional layer to perform primary feature extraction and a first dimensionality reduction operation to determine a first eigenvector; The first feature vector is sequentially input into the first activation function layer, the first activation function layer and the average pooling layer to obtain the first feature data after the second dimensionality reduction operation.

4. The method according to claim 2, characterized in that The deep feature extraction module includes at least three attention mechanism layers, a second activation function layer, and a maximum pooling layer. The first feature data is input into the deep feature extraction module of the initial state monitoring model to obtain the second feature data, including: Inputting the first feature data into the at least three attention mechanism layers in sequence to determine the processed hidden feature information and key feature information; The hidden feature information and the key feature information are sequentially input into the second activation function layer and the maximum pooling layer to obtain second feature data.

5. The method according to claim 4, characterized in that Inputting the first feature data into the at least three attention mechanism layers in sequence, and determining the processed hidden feature information and key feature information, including: The first feature data is sequentially input into the at least three attention mechanism layers, and the processed hidden feature information and key feature information are determined by a first formula, wherein the first formula is expressed as: Where X represents a vector sequence of N elements in the first feature data; q represents a query vector; α n represents the weight of the nth element, and 6. The method according to claim 2, characterized in that Inputting the second characteristic data into the mapping relationship building module of the initial state monitoring model to determine and build a nonlinear mapping relationship between the vibration signal of the rotating machinery and the health state includes: Inputting the second characteristic data into the temporal convolutional neural network of the mapping relationship building module to determine and build a nonlinear mapping relationship between the vibration signal of the rotating machinery and the health status; The temporal convolutional neural network comprises at least: an expanded causal convolution layer, a weight normalization layer, a clipping function layer, a third activation function layer and a random inactivation dropout layer.

7. The method according to claim 1, characterized in that The vibration signal is subjected to signal conversion and noise reduction processing to obtain an experimental data set, including: Performing signal conversion on the vibration signal to determine a digital signal of the vibration signal; performing time domain characteristic analysis, data defect preprocessing and filling, and data noise reduction processing on the digital signal to determine processed first data; the time domain characteristic analysis includes calculating at least one time domain characteristic of the digital signal including an average value, variance, standard deviation, and peak value, performing feature extraction with time as a variable, and determining a characteristic of the digital signal changing over time; The first data is formatted in a unified manner to determine a plurality of groups of processed data samples, wherein the plurality of groups of data samples have a unified dimension and length; Each group of data samples is assigned a data label to obtain an experimental data set.

8. A rotating machinery state monitoring device, characterized in that: include: An acquisition module is used to acquire vibration signals during the operation of the rotating machinery; A first determination module is configured to perform signal conversion and noise reduction processing on the vibration signal to obtain an experimental data set; A second determination module is configured to train a target rotating machinery state monitoring model based on the experimental data set; The first processing module is configured to monitor the state of the rotating machinery using the target rotating machinery state monitoring model.

9. A readable storage medium having a program or instruction stored thereon, characterized in that: When the program or instruction is executed by a processor, the steps in the method according to any one of claims 1 to 7 are implemented.

10. A computer program product, characterized in that The method comprises computer instructions, which, when executed by a processor, implement the steps of the method according to any one of claims 1 to 7.

Images