A method, device, and storage medium for identifying mechanical equipment failures

Abstract

The present invention belongs to the technical field of mechanical intelligent operation and maintenance, and discloses a method, device and storage medium for mechanical equipment fault identification in the technical field of mechanical equipment intelligent operation and maintenance. The method includes: extracting the domain global features of the vibration signal, extracting the in-domain classification features of the domain global features through the in-domain classification feature extractor, extracting the inter-domain classification features of the vibration signal through the inter-domain feature extractor, performing a numerical addition operation on the inter-domain classification features and the in-domain classification features to obtain the recombined classification features, and inputting the inter-domain classification features, the in-domain classification features and the recombined classification features into a fault classifier to obtain the fault type prediction probability; obtaining the corresponding fault type ordinal number through the maximum value of the fault type prediction probability, so as to realize the identification of the fault type. The present invention can solve the technical problem that new faults are misidentified as existing fault types, resulting in a decrease in the accuracy of mechanical equipment fault identification.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical intelligent operation and maintenance, and particularly relates to a method, device and storage medium for identifying mechanical equipment faults. Background Art

[0002] With the great development of modern mechanical systems, the need for their prediction and health management has become increasingly urgent. During the operation of mechanical systems, the occurrence of mechanical faults will seriously affect the safety and efficiency of industrial production.

[0003] In the era of industrial big data, the traditional method of inspecting mechanical equipment is time-consuming and laborious, and it has been difficult to meet the accurate identification requirements for mechanical equipment faults. The intelligent maintenance of machinery has become a research hotspot in recent years. Due to its outstanding automation ability, the mechanical equipment fault diagnosis technology based on deep learning can better ensure the safety and reliability of mechanical equipment, especially mechanical equipment parts such as rolling bearings and gears under harsh operating conditions during operation, providing a new feasible solution for mechanical equipment fault diagnosis.

[0004] However, the mechanical equipment fault diagnosis technology based on deep learning is restricted in two aspects in practical engineering applications. On the one hand, the operating conditions of mechanical equipment are variable, and the state data of mechanical equipment collected under different conditions have different distributions, and the distribution differences pose challenges to the extraction of mechanical equipment state characteristics. On the other hand, the collected historical loss type data cannot cover all fault types. When the model performs online identification, new types of faults encountered will be misidentified as existing fault types, resulting in a decrease in the accuracy of mechanical equipment fault identification by deep learning, and it is difficult to guarantee the safety and efficiency of mechanical equipment operation.

[0005] Therefore, there is an urgent need for a method, device and storage medium for identifying mechanical equipment faults to solve the above technical problems. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method, device and storage medium for identifying mechanical equipment faults, which can solve the technical problem that new types of faults will be misidentified as existing fault types, resulting in a decrease in the accuracy of mechanical equipment fault identification.

[0007] To achieve the above object, the present invention is implemented by the following technical solutions:

[0008] In the first aspect, the present invention provides a method for identifying mechanical equipment faults, including:

[0009] Obtain the vibration signal of the mechanical equipment;

[0010] Input the vibration signal into a pre-trained diagnostic model, where the diagnostic model includes a domain global feature extractor, an inter-domain feature extractor, an intra-domain classification feature extractor, and a fault classifier;

[0011] Extract the domain global features of the vibration signal through the domain global feature extractor, extract the intra-domain classification features of the domain global features through the intra-domain classification feature extractor, extract the inter-domain classification features of the vibration signal through the inter-domain feature extractor, perform a numerical addition operation on the inter-domain classification features and the intra-domain classification features to obtain recombined classification features, and input the inter-domain classification features, intra-domain classification features, and recombined classification features into the fault classifier to obtain the fault type prediction probability;

[0012] Obtain the corresponding fault type ordinal number through the maximum value of the fault type prediction probability to realize the identification of the fault type.

[0013] Furthermore, the training process of the diagnostic model includes:

[0014] Perform non-overlapping segmentation on the collected historical vibration signals according to a preset segmentation length to obtain state samples, and calibrate according to the fault types and the types of the acquisition working condition domains of the state samples to obtain a labeled training data set;

[0015] Extract the domain global features of the state samples in the labeled training data set through the domain global feature extractor, input the domain global features into the intra-domain classification feature extractor to extract the intra-domain classification features of the state samples, and extract the inter-domain classification features of the state samples through the inter-domain feature extractor;

[0016] Perform a numerical addition operation on the inter-domain classification features and the intra-domain classification features of the state samples to obtain the recombined classification features of the state samples;

[0017] Input the inter-domain classification features, intra-domain classification features, and recombined classification features of the state samples into the fault classifier to obtain the fault type prediction probability of the state samples;

[0018] Calculate the fault classification cross-entropy loss according to the fault type prediction probability of the state samples and the true fault type labels calibrated for the state samples, calculate the preference correction loss based on the fault type prediction probability, minimize the fault classification cross-entropy loss and the preference correction loss according to the backpropagation and optimization algorithms, and determine the trained diagnostic model according to the minimum fault classification cross-entropy loss and preference correction loss.

[0019] Furthermore, the diagnostic model further includes a working condition specific feature extractor and a domain classifier, and the training process of the diagnostic model further includes:

[0020] Input the domain global features of the state sample into the working condition specific feature extractor to extract the working condition specific features, and input the domain global features, working condition specific features and in-domain classification features of the state sample into the domain classifier to obtain the predicted probability of the domain type of the state sample;

[0021] Calculate the domain classification cross-entropy loss according to the predicted probability of the domain type and the true domain type label in the labeled training dataset, calculate the covariance loss according to the working condition specific features, in-domain classification features and inter-domain classification features, minimize the domain classification cross-entropy loss and covariance loss based on backpropagation and optimization algorithms, determine the optimal training results of the working condition specific feature extractor and the domain classifier according to the minimum domain classification cross-entropy loss and covariance loss, and take the optimal training results of the working condition specific feature extractor and the domain classifier as the finally trained diagnostic model to complete the training.

[0022] Further, calculating the fault classification cross-entropy loss according to the predicted probability of the fault type and the true fault type label in the labeled training dataset includes:

[0023] ,

[0024] ,

[0025] where, is the fault classification cross-entropy loss, represents the sample ordinal number, represents the total number of samples, represents the one-hot vector of the true domain type label of the th state sample, represents the predicted probability of the fault type of the th state sample, represents the predicted probability that the th state sample belongs to the th fault type, K represents the total number of known fault types in the labeled training dataset, represents the new fault type.

[0026] Further, calculating the preference correction loss based on the predicted probability of the fault type includes:

[0027] ,

[0028] where, is the preference correction loss, represents the sample ordinal number, represents the total number of samples, represents the predicted probability that the th state sample belongs to the th fault type, Indicates the ordinal number of the fault type corresponding to the th state sample, Indicates the predicted probability that the th state sample belongs to a new fault type, Indicates the new fault type.

[0029] Further, calculating the domain classification cross-entropy loss according to the domain type prediction probability and the true domain type label calibrated by the state sample includes:

[0030] ,

[0031] ,

[0032] wherein, is the domain classification cross-entropy loss, Indicates the sample ordinal number, Indicates the total number of samples. In the formula, Indicates the one-hot vector of the true domain type label of the th state sample, Indicates the domain type prediction probability of the th state sample, Indicates the predicted probability that the th state sample belongs to the th domain type, and M is the total number of domains;

[0033] Calculating the covariance loss according to the working condition specific features, intra-domain classification features and inter-domain classification features includes:

[0034] ,

[0035] wherein, is the covariance loss, Indicates the working condition specific features, Indicates the intra-domain classification features, Indicates the inter-domain classification features, Indicates the covariance matrix, Indicates the Euclidean norm.

[0036] Further, the optimization algorithm includes one of the adaptive moment estimation algorithm, the stochastic gradient descent algorithm, and the root mean square propagation algorithm.

[0037] In a second aspect, the present invention provides a mechanical equipment fault identification device, including:

[0038] An acquisition module for acquiring the vibration signal of the mechanical equipment;

[0039] An input module for inputting the vibration signal into a pre-trained diagnostic model, where the diagnostic model includes a domain global feature extractor, an inter-domain feature extractor, an intra-domain classification feature extractor, and a fault classifier;

[0040] A prediction module for extracting the domain global features of the vibration signal through the domain global feature extractor, extracting the intra-domain classification features of the domain global features through the intra-domain classification feature extractor, extracting the inter-domain classification features of the vibration signal through the inter-domain feature extractor, performing a numerical addition operation on the inter-domain classification features and the intra-domain classification features to obtain recombined classification features, and inputting the inter-domain classification features, the intra-domain classification features, and the recombined classification features into the fault classifier to obtain the fault type prediction probability;

[0041] An identification module for obtaining the corresponding fault type ordinal number through the maximum value of the fault type prediction probability to achieve the identification of the fault type;

[0042] Among them, the multiple fault types of the vibration signal include the fault types of the training samples and the new fault types, and the fault type ordinal number is greater than the total number of the fault types of the training samples.

[0043] In a third aspect, the present invention provides an electronic terminal, including a processor and a memory connected to the processor, where a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the method described in any one of the above are executed.

[0044] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method described in any one of the above are implemented.

[0045] Beneficial effects

[0046] In the present invention, by inputting the vibration signal to be identified into a pre-constructed diagnostic model, since the fault classifier in the diagnostic model has more set identification numbers, the fault types of the training samples and the new fault types can be identified, and the fault type prediction probabilities corresponding to multiple fault types are obtained. Among them, the multiple fault types include the fault types of the training samples and the new fault types, realizing the identification of the new fault types, improving the accuracy of mechanical equipment fault identification, and ensuring the safety and efficiency of mechanical equipment operation;

[0047] The fault classifier will have K + 1 prediction probabilities for all training samples. However, since there are no new fault samples participating in the training, the prediction preference of the classifier will tend to the known fault samples. Therefore, by setting the preference correction loss, the classifier preference will be corrected to balance the prediction probabilities of the known fault types and the new fault types;

[0048] In the diagnostic model training stage, the domain classification loss is calculated through the domain type prediction probability to optimize the domain-related features, namely, the domain global features, the working condition specific features, and the in-domain classification features. Also, the covariance loss is calculated based on the working condition specific features, the in-domain classification features, and the inter-domain classification features to further decouple the features, ensure the mutual exclusion of information between different features, facilitate the recombination of the fault classification features, and extract more generalizable fault features. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 FIG. is a schematic flowchart of a mechanical equipment fault identification method provided in Embodiment 1 of the present invention;

[0050] Figure 2 FIG. is a schematic structural diagram of a diagnostic model of a mechanical equipment fault identification method provided in Embodiment 1 of the present invention;

[0051] Figure 3 FIG. is a schematic diagram of on-line identification of a vibration signal to be identified in a mechanical equipment fault identification method provided in Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] The technical solution of the present invention will be described in detail below with reference to the drawings and specific embodiments. It should be understood that the specific features in the embodiments of the present application and the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application. Without conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other.

[0053] The term "and / or" in this article is merely a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this article generally represents an "or" relationship between the front and rear associated objects.

[0054] Non-overlapping segmentation: Non-overlapping segmentation of vibration signals generally means that in the signal processing process, the signal is divided into multiple continuous segments without overlapping parts between each segment. It allows for higher resolution in the frequency domain while reducing signal distortion caused by window functions;

[0055] In practical applications, which method to choose depends on the characteristics of the signal and the analysis objectives. If non-overlapping segmentation is required, the maximum overlap discrete wavelet transform (MODWT) can be considered. These methods can process signals of any length and can avoid a series of problems caused by overlap.

[0056] Regarding how to "obtain the fault type prediction probability corresponding to multiple fault types through the fault classifier based on the intra-domain classification features and inter-domain classification features", those skilled in the art can understand that the "fault classifier" in this application can obtain the specific fault type corresponding to the data to be tested under its corresponding classification logic just like a conventional neural network model classifier after training according to the optimization training process proposed in this application; "inputting the domain global features, working condition specific features and intra-domain classification features into the domain classifier to obtain the domain type prediction probability of the state sample" is the same and will not be repeated.

[0057] Example 1

[0058] Figure 1 This is a flow chart of the mechanical equipment fault identification method in the first embodiment of the present invention. This flow chart only shows the logical sequence of the method described in this embodiment. In other possible embodiments of the present invention, different Figure 1 The steps shown or described are accomplished in the order shown.

[0059] The mechanical equipment fault identification method provided in this embodiment can be applied to a terminal and can be executed by a mechanical equipment fault identification device, which can be implemented by software and / or hardware, and can be integrated in a terminal, for example: any smart phone, tablet computer or computer device with communication function. Figure 1 and Figure 3 As shown, the method of this embodiment specifically includes the following steps:

[0060] Step 1: Obtain different types of vibration signals under historical working conditions and target working conditions of mechanical equipment.

[0061] Step 2: Input the vibration signal into a pre-trained diagnostic model, such as Figure 2 As shown, the diagnostic model includes a domain global feature extractor, an inter-domain feature extractor, an intra-domain classification feature extractor, a working condition specific feature extractor, a domain classifier and a fault classifier. Before the vibration signal is input into the pre-trained diagnostic model, the collected historical vibration signal can be non-overlappingly segmented according to a preset segmentation length to obtain a state sample, which is calibrated according to the fault type of the state sample and the acquisition working condition domain type to obtain an unlabeled test data set, and finally the unlabeled test data set including the vibration signal to be identified is input into the diagnostic model.

[0062] Step 3: Extract the domain-global features of the vibration signal through the domain-global feature extractor, extract the in-domain classification features of the domain-global features through the in-domain classification feature extractor, extract the inter-domain classification features of the vibration signal through the inter-domain feature extractor. Based on the in-domain classification features and inter-domain classification features, it should be noted that in the test phase, the input of the fault classifier is as Figure 3 shown, which is only the in-domain classification features and inter-domain classification features, and obtain the fault type prediction probabilities corresponding to multiple fault types through the fault classifier;

[0063] Step 4: Obtain the corresponding fault type ordinal number through the maximum value of the fault type prediction probability to realize the identification of the fault type;

[0064] Among them, the multiple fault types of the vibration signal include the fault types of the training samples and new fault types, and the fault type ordinal number is greater than the total number of fault types of the training samples.

[0065] Specifically, the training process of the diagnosis model includes:

[0066] Perform non-overlapping segmentation on the collected historical vibration signals according to a preset segmentation length to obtain state samples, and calibrate according to the fault types and acquisition working condition domain types of the state samples to obtain a labeled training data set;

[0067] Extract the domain-global features of the state samples in the training data set through the domain-global feature extractor, input the domain-global features into the in-domain classification feature extractor to extract the in-domain classification features of the state samples, and extract the inter-domain classification features of the state samples through the inter-domain feature extractor;

[0068] Perform a numerical addition operation on the inter-domain classification features and in-domain classification features of the state samples to obtain the recombined classification features of the state samples;

[0069] Input the inter-domain classification features, in-domain classification features, and recombined classification features of the state samples into the fault classifier to obtain the fault type prediction probabilities of the state samples;

[0070] Calculate the fault classification cross-entropy loss according to the fault type prediction probabilities and the true fault type labels in the labeled training data set, including:

[0071] ,

[0072] ,

[0073] Among them, is the fault classification cross-entropy loss, represents the sample ordinal number, represents the total number of samples, Indicates the one - hot vector of the real - domain type label of the th state sample, Indicates the predicted probability of the fault type of the th state sample, Indicates the predicted probability that the th state sample belongs to the th fault type. K represents the total number of known fault types in the labeled training dataset, Indicates the new fault type;

[0074] Calculating the preference - correction loss based on the predicted probability of the fault type includes:

[0075] ,

[0076] where, is the preference - correction loss, represents the sample ordinal number, represents the total number of samples, Indicates the predicted probability that the th state sample belongs to the th fault type, Indicates the ordinal number of the fault type corresponding to the th state sample, Indicates the predicted probability that the th state sample belongs to the new fault type, Indicates the new fault type.

[0077] Calculate the fault - classification cross - entropy loss according to the predicted probability of the fault type of the state sample and the true fault - type label calibrated for the state sample, calculate the preference - correction loss based on the predicted probability of the fault type, minimize the fault - classification cross - entropy loss and the preference - correction loss according to the back - propagation and optimization algorithms, and determine the trained diagnostic model according to the minimum fault - classification cross - entropy loss and preference - correction loss.

[0078] Regarding the preference - correction loss: During the training process, the fault classifier generates K + 1 predicted probabilities for each training sample (corresponding to K known fault types and the K + 1th new fault type, and the specific new fault type is unknown). However, since there are no new fault samples participating in the training during the training process, the prediction preference of the fault classifier will be biased towards the known fault types (the predicted probabilities for the K known fault types will be greater than the predicted probability for the K + 1th new fault type). At this time, the preference - correction loss will correct the classifier preference to balance the predicted probabilities of the fault classifier for known and unknown faults.

[0079] Specifically, the training process of the diagnostic model further includes:

[0080] Input the domain global features of the state sample into the working condition specific feature extractor to extract the working condition specific features. Input the domain global features, the working condition specific features and the within-domain classification features of the state sample into the domain classifier to obtain the domain type prediction probability of the state sample;

[0081] Calculating the domain classification cross-entropy loss according to the domain type prediction probability and the true domain type label in the labeled training dataset includes:

[0082] ,

[0083] ,

[0084] where, is the domain classification cross-entropy loss, represents the sample ordinal number, represents the total number of samples. In the formula, represents the one-hot vector of the true domain type label of the th state sample, represents the domain type prediction probability of the th state sample, represents the prediction probability that the th state sample belongs to the th domain type, and M is the total number of domains;

[0085] Calculating the covariance loss according to the working condition specific features, the within-domain classification features and the between-domain classification features includes:

[0086] ,

[0087] where, is the covariance loss, represents the working condition specific features, represents the within-domain classification features, represents the between-domain classification features, represents the covariance matrix, represents the Euclidean norm;

[0088] Based on backpropagation and optimization algorithms, minimize the domain classification cross-entropy loss and the covariance loss. Determine the optimal training results of the working condition specific feature extractor and the domain classifier according to the minimum domain classification cross-entropy loss and covariance loss. Take the optimal training results of the working condition specific feature extractor and the domain classifier as the finally trained diagnostic model to complete the training.

[0089] It should be noted that the predicted probability of the computational domain type here is for calculating the domain cross-entropy loss, and its purpose is to optimize the working condition specific features, the in-domain classification features, and the domain global features extracted by the domain global feature extractor. The covariance loss is calculated and minimized based on the domain global features, the in-domain classification features, and the working condition specific features to achieve feature decoupling. Finally, the information contained in each feature is made mutually exclusive, so that each feature vector contains unique information.

[0090] The optimization algorithm adopts one of the adaptive moment estimation algorithm, the stochastic gradient descent algorithm, and the root mean square propagation algorithm. The backpropagation refers to calculating the gradient of the obtained loss, and then the optimization algorithm updates the network parameters according to the obtained gradient. They all belong to the prior art and will not be elaborated here.

[0091] The following further describes this embodiment in conjunction with the drawings and experimental cases:

[0092] 1. Experimental data

[0093] The data used in the experiment comes from the open-source dataset and the self-made dataset in the laboratory. As shown in Table 1, the open-source dataset uses an accelerometer to collect the vibration signals of artificially damaged bearings through a simulation test bench, and the sampling frequency is 25.6 kHz. This embodiment considers five bearing health states: normal state, inner race fault, outer race fault, roller fault, and a compound fault of inner race and outer race fault. Their corresponding classification labels are 0, 1, 2, 3, and 4 respectively. At the same time, three different rotational speed conditions are simulated, and 200 sample data are taken for each of the five bearing health states at three rotational speeds of 4200 (r / min), 4500 (r / min), and 4800 (r / min) respectively, and are set as different domains H1, H2, and H3. The self-made dataset in the laboratory installs an accelerometer on the bearing housing through a simulation test bench to capture vibration data at a sampling rate of 32768 Hz. In this case, the six health conditions including labels are as follows: normal state, inner race fault, outer race fault, and three compound faults: inner race and roller fault, inner race and outer race fault, outer race and roller fault. Their corresponding class labels are 0, 1, 2, 3, 4, and 5 respectively. At the same time, 4 different working conditions are simulated, and the rotational speed and radial force in each working condition are different. 200 sample data are taken for each of the six bearing health states in each working condition, and are set as different domains S1, S2, S3, and S4. Each state sample of the above two datasets contains 1024 data points.

[0094] Table 1 Open-source dataset and self-made dataset in the laboratory

[0095]

[0096] 2. Method verification

[0097] To verify the effectiveness of the proposed invention, in this example, eight tasks were constructed using the above two datasets for experimental verification. As shown in Table 2, the training set data and target working condition data of eight tasks under two datasets were constructed by matching different fault types and different working condition conditions. Among them, task number one uses an open-source dataset to construct the task, and task number two uses a self-made dataset in the laboratory to construct the task. In this embodiment, multiple diagnostic tasks are set. The fault types that are not included in the training set labels but are included in the target working condition labels in each diagnostic task are new faults. For example, for task T1 in Table 2, the inner and outer ring faults represented by label 4 are new faults. To further verify the advancement of the invention, we simultaneously used several other fault detection methods in this field and the invention to execute these tasks for comparison, and the H-index was used to represent the fault diagnosis ability of the method. Among them, the H-index is defined as the harmonic mean of the classification accuracy of known fault classes and the classification accuracy of unknown fault classes, and it is a commonly used indicator in open-set problems. Only when both classification accuracies are good, it shows a high value. The experimental results are shown in Table 3, where M1 is the adaptive open-set domain generalization network model, M2 is the meta-learning model for joint domain class matching, M3 is the extreme value theory model based on convolutional neural network, and M4 is the model of the invention based on empirical threshold. Among the relevant parameters of the model of the invention, the training period and batch size are set to 150 and 10 respectively, and the learning rate is set to 0.01.

[0098] Table 2 Task settings for two datasets

[0099]

[0100] Table 3 Comparative experiment of different methods, histochemical score (H-score) and standard deviation

[0101]

[0102] Comparing the data in Table 3, therefore, we can draw the following conclusions: 1) On the open-source dataset and the self-made dataset in the laboratory, the average H-index of M4 and the method proposed in this invention is mostly higher than that of M1, M2, and M3. This shows the effectiveness of the invention and the advancement of the idea of feature recombination and open-set classifier preference correction in diagnostic tasks containing unknown fault types; 2) The average H-index of this invention on the open-source dataset and the self-made dataset in the laboratory exceeds the M4 category by 18.3% and 8.1% respectively, which respectively proves the superiority of the designed classification preference correction loss in learning the decision boundary; 3) The average H-values of this invention on the open-source dataset and the self-made dataset in the laboratory are 83.6% and 80.4% respectively, showing excellent performance in all tasks. The results verify the superior classification ability of this invention for known and unknown fault classes under unknown working conditions.

[0103] Example 2

[0104] Embodiment 2 of the present invention provides a mechanical equipment fault identification device, including:

[0105] An acquisition module, configured to acquire vibration signals of mechanical equipment;

[0106] An input module, configured to input the vibration signals into a pre-trained diagnostic model, where the diagnostic model includes a domain global feature extractor, an inter-domain feature extractor, an intra-domain classification feature extractor, and a fault classifier;

[0107] A prediction module, configured to extract the domain global features of the vibration signals through the domain global feature extractor, extract the intra-domain classification features of the domain global features through the intra-domain classification feature extractor, extract the inter-domain classification features of the vibration signals through the inter-domain feature extractor, and obtain the fault type prediction probabilities corresponding to multiple fault types through the fault classifier based on the intra-domain classification features and the inter-domain classification features;

[0108] An identification module, configured to obtain the corresponding fault type ordinal number through the maximum value of the fault type prediction probabilities, so as to realize the identification of the fault type;

[0109] Wherein, the multiple fault types of the vibration signals include the fault types of training samples and new fault types, and the fault type ordinal number is greater than the total number of fault types of training samples.

[0110] The mechanical equipment fault identification device provided in Embodiment 2 of the present invention can execute the mechanical equipment fault identification method provided in Embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects for executing the method.

[0111] Example 3

[0112] Embodiment 3 of the present invention further provides an electronic terminal, including a processor and a memory connected to the processor, where a computer program is stored in the memory, and the processor is configured to operate according to the instructions to execute the steps of the method in Embodiment 1.

[0113] The electronic terminal provided in Embodiment 3 of the present invention can execute the mechanical equipment fault identification method provided in Embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects for executing the method.

[0114] Example 4

[0115] Embodiment 4 of the present invention further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it realizes the steps of the method in Embodiment 1, and has the corresponding functional modules and beneficial effects for executing the method.

[0116] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, an apparatus, or a computer program product. Therefore, the present application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0117] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (apparatuses), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the flows and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0118] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Therefore, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0120] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for identifying mechanical equipment faults, characterized in that: include: Obtain vibration signals from mechanical equipment; Inputting the vibration signal into a pre-trained diagnostic model, wherein the diagnostic model includes a domain global feature extractor, an inter-domain feature extractor, an intra-domain classification feature extractor, and a fault classifier; The domain global feature extractor is used to extract the domain global feature of the vibration signal, the intra-domain classification feature extractor is used to extract the intra-domain classification feature of the domain global feature, the inter-domain classification feature of the vibration signal is extracted by the inter-domain feature extractor, the inter-domain classification feature and the intra-domain classification feature are numerically added to obtain the recombined classification feature, and the inter-domain classification feature, the intra-domain classification feature and the recombined classification feature are input into the fault classifier to obtain the prediction probability of the fault type; The corresponding fault type ordinal number is obtained by the maximum value of the predicted probability of the fault type, so as to realize the identification of the fault type; The training process of the diagnostic model includes: Perform non-overlap segmentation on the collected historical vibration signals according to a preset segmentation length to obtain state samples, calibrate the state samples according to the fault type and the collected working condition domain type, and obtain a labeled training data set; Extracting domain global features of state samples in the labeled training data set by the domain global feature extractor, inputting the domain global features into the intra-domain classification feature extractor to extract intra-domain classification features of the state samples, and extracting inter-domain classification features of the state samples by the inter-domain feature extractor; Performing numerical addition operation on the inter-domain classification features and the intra-domain classification features of the state sample to obtain the reorganization classification features of the state sample; inputting the inter-domain classification features, the intra-domain classification features and the reorganization classification features of the state sample into a fault classifier to obtain the prediction probability of the fault type of the state sample; Calculate the fault classification cross entropy loss according to the fault type prediction probability of the state sample and the real fault type label calibrated by the state sample, calculate the preference correction loss based on the fault type prediction probability, minimize the fault classification cross entropy loss and the preference correction loss according to the back propagation and optimization algorithm, and determine the trained diagnosis model according to the minimum fault classification cross entropy loss and the preference correction loss; The domain global features of the state sample are input into the working condition specific feature extractor to extract the working condition specific features, and the domain global features, working condition specific features and intra-domain classification features of the state sample are input into the domain classifier to obtain the domain type prediction probability of the state sample.

2. The mechanical equipment fault identification method according to claim 1, characterized in that: The diagnostic model further includes a condition-specific feature extractor and a domain classifier, and the training process of the diagnostic model further includes: The domain classification cross entropy loss is calculated according to the domain type prediction probability and the real domain type label calibrated by the state sample, the covariance loss is calculated according to the working condition specific features, the intra-domain classification features and the inter-domain classification features, the domain classification cross entropy loss and the covariance loss are minimized based on the back propagation and optimization algorithm, the optimal training results of the working condition specific feature extractor and the domain classifier are determined according to the minimum domain classification cross entropy loss and the covariance loss, and the optimal training results of the working condition specific feature extractor and the domain classifier are used as the working condition specific feature extractor and the domain classifier of the finally trained diagnostic model.

3. The mechanical equipment fault identification method according to claim 1, characterized in that: The fault classification cross entropy loss is calculated according to the predicted probability of the fault type of the state sample and the real fault type label of the state sample, including: ， ， in, is the cross entropy loss for fault classification, represents the sample ordinal number, represents the total number of samples, Indicates The one-hot vector of the true domain type label of each state sample, Indicates The predicted probability of the fault type of the state sample is Indicates The state samples belong to The predicted probability of each failure type, K represents the total number of known fault types in the labeled training dataset, Indicates a new fault type.

4. The mechanical equipment fault identification method according to claim 1, characterized in that: Calculating the preference correction loss based on the predicted probability of the fault type includes: ， in, is the preference correction loss, represents the sample ordinal number, represents the total number of samples, Indicates The state samples belong to The predicted probability of each failure type, Indicates The fault type ordinal number corresponding to the state sample, Indicates The predicted probability that a state sample belongs to a new fault type, Indicates a new fault type.

5. The mechanical equipment fault identification method according to claim 2, characterized in that: Calculating the domain classification cross entropy loss based on the domain type prediction probability and the true domain type label of the state sample includes: ， ， in, is the domain classification cross entropy loss, represents the sample ordinal number, represents the total number of samples, where Indicates The one-hot vector of the true domain type label of each state sample, Indicates The domain type prediction probability of state samples, Indicates The state samples belong to The predicted probability of each domain type, M is the total number of domains; Calculating the covariance loss based on the working condition specific features, the intra-domain classification features, and the inter-domain classification features includes: ， in, is the covariance loss, Indicates the specific characteristics of the working condition, represents the classification features within the domain, represents the inter-domain classification feature, represents the covariance matrix, represents the Euclidean norm.

6. The mechanical equipment fault identification method according to claim 1, characterized in that: The optimization algorithm includes one of an adaptive moment estimation algorithm, a stochastic gradient descent algorithm and a root mean square transfer algorithm.

7. A mechanical equipment fault identification device, characterized in that: include: Acquisition module, used to obtain vibration signals of mechanical equipment; An input module, used for inputting the vibration signal into a pre-trained diagnostic model, wherein the diagnostic model includes a domain global feature extractor, an inter-domain feature extractor, an intra-domain classification feature extractor, and a fault classifier; A prediction module, configured to extract a domain global feature of the vibration signal through the domain global feature extractor, extract an intra-domain classification feature of the domain global feature through the intra-domain classification feature extractor, extract an inter-domain classification feature of the vibration signal through the inter-domain feature extractor, perform a numerical addition operation on the inter-domain classification feature and the intra-domain classification feature to obtain a recombined classification feature, and input the inter-domain classification feature, the intra-domain classification feature and the recombined classification feature into a fault classifier to obtain a prediction probability of a fault type; An identification module, used to obtain a corresponding fault type ordinal number through the maximum value of the predicted probability of the fault type, so as to identify the fault type; The multiple fault types of the vibration signal include fault types of training samples and new fault types, and the fault type ordinal number is greater than the total number of fault types of training samples; The training process of the diagnostic model includes: Perform non-overlap segmentation on the collected historical vibration signals according to a preset segmentation length to obtain state samples, calibrate the state samples according to the fault type and the collected working condition domain type, and obtain a labeled training data set; Extracting domain global features of state samples in the labeled training data set by the domain global feature extractor, inputting the domain global features into the intra-domain classification feature extractor to extract intra-domain classification features of the state samples, and extracting inter-domain classification features of the state samples by the inter-domain feature extractor; Performing numerical addition operation on the inter-domain classification features and the intra-domain classification features of the state sample to obtain the reorganization classification features of the state sample; inputting the inter-domain classification features, the intra-domain classification features and the reorganization classification features of the state sample into a fault classifier to obtain the prediction probability of the fault type of the state sample; Calculate the fault classification cross entropy loss according to the fault type prediction probability of the state sample and the real fault type label calibrated by the state sample, calculate the preference correction loss based on the fault type prediction probability, minimize the fault classification cross entropy loss and the preference correction loss according to the back propagation and optimization algorithm, and determine the trained diagnosis model according to the minimum fault classification cross entropy loss and the preference correction loss; The domain global features of the state sample are input into the working condition specific feature extractor to extract the working condition specific features, and the domain global features, working condition specific features and intra-domain classification features of the state sample are input into the domain classifier to obtain the domain type prediction probability of the state sample.

8. An electronic terminal, characterized in that: The method comprises a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 6 are executed.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images