A mechanical equipment fault diagnosis method, device, equipment and readable storage medium

Abstract

The application discloses a mechanical equipment fault diagnosis method, device and equipment and a readable storage medium. Known fault information multi-source frequency domain signals of same or same type mechanical equipment under a first working condition are acquired, and the known fault information multi-source frequency domain signals under the first working condition are taken as source domain data sets. Unknown fault information multi-source frequency domain signals of the mechanical equipment to be diagnosed under a second working condition are acquired, and the unknown fault information multi-source frequency domain signals under the second working condition are taken as target domain data sets. A multi-source dense adaptive adversarial network model is trained by using the source domain data sets and the target domain data sets, and a mechanical equipment fault diagnosis model is obtained. The real-time acquired multi-source frequency domain signals of the mechanical equipment to be diagnosed under the second working condition are input into the mechanical equipment fault diagnosis model, and a fault diagnosis result is output. The application can better utilize the multi-source information of the mechanical equipment in data level fusion and feature level fusion.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment fault diagnosis technology, specifically relating to a mechanical equipment fault diagnosis method, device, equipment, and readable storage medium. Background Technology

[0002] With the rapid development of modern industry, machinery and equipment have become the cornerstone of improving productivity and promoting economic growth. The mechanical transmission systems of rotating machinery operate in harsh environments for extended periods, inevitably leading to malfunctions that can cause accidents and economic losses. To improve the reliability of machinery and equipment and reduce maintenance costs, numerous advanced signal processing technologies and data-driven methods have been developed for the diagnosis of rotating machinery and equipment.

[0003] In recent years, deep learning (DL) has been widely applied in the field of mechanical equipment fault diagnosis due to its powerful feature extraction and pattern recognition capabilities. LONG et al. proposed a hybrid intelligent diagnostic model based on sparse autoencoders and support vector machines for the posture data of multi-joint industrial robots. Furthermore, some researchers have proposed a deep convolutional neural network to convolve the original multi-channel data, achieving ideal recognition results. LI et al. introduced a hybrid evolutionary algorithm and hierarchical optimization strategy into echo state networks, demonstrating good performance. However, classic DL methods require a fundamental assumption: that training and test data are independent and have the same distribution. Unfortunately, in practical industrial applications, data often comes from different working conditions, such as different speeds or loads, all affected by environmental changes. Since this data does not meet the above assumption, DL methods may fail to meet fault diagnosis requirements due to performance degradation. Moreover, it is difficult to obtain sufficient labeling information for test samples in industrial settings. To reliably diagnose these fault data, the trained model must be modified to adapt to the distribution of the target samples. Unfortunately, this process requires significant time and resources.

[0004] To address the aforementioned issues, domain adaptation has become an important sub-direction within transfer learning (TL), effectively solving practical problems. In recent years, many scholars have developed various domain adaptation methods in the field of mechanical equipment fault diagnosis. By simultaneously considering the completeness, interpretability, and transferability of features, these methods solve cross-domain diagnosis and knowledge-based transfer problems under different operating conditions. Extending marginal distribution domain adaptation to joint distribution domain adaptation further improves the accuracy of deep network-based fault diagnosis methods. The aforementioned research focuses on reducing metric statistics between different domains. Recently, the idea of ​​adversarial learning has also been applied to domain adaptation, achieving fault knowledge transfer through adversarial interaction between discriminators and classifiers.

[0005] Despite significant progress and breakthroughs in domain adaptation research in fault diagnosis, the aforementioned research still has the following shortcomings. First, it fails to adequately utilize multi-source information, particularly in data-level and feature-level fusion. Sensors deployed on actual devices always contain multiple sensors and even various types of multi-mode signals. Current research is mostly based on single-sensor signals, failing to fully utilize multi-source information. Second, in feature extraction using deep models, existing methods simply splice and combine features, making it difficult to accurately characterize the machine's health status. Third, numerous domain adaptation methods have been proposed during training to reduce the feature distribution differences of the proposed networks. However, these works do not address the inherent label information in fault classification problems in detail. In fact, the distribution difference between the labels of source domain samples and the pseudo-labels of target domain samples generated by the network can also serve as an optimization objective. Furthermore, during the training process of transfer tasks, transferring abnormal samples presents difficulties, negatively impacting the overall domain adaptation and resulting in unsatisfactory transfer performance. This places higher demands on domain adaptation in fault diagnosis. Summary of the Invention

[0006] To address the problems existing in the prior art, the present invention provides a method, apparatus, device and readable storage medium for diagnosing mechanical equipment faults. It can better utilize multi-source information of mechanical equipment in data-level fusion and feature-level fusion, has a stronger feature extraction capability, and can effectively utilize the inherent label information in the fault classification problem and more effectively transfer abnormal samples in the process of domain adaptation.

[0007] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0008] A method for diagnosing mechanical equipment faults, comprising:

[0009] Acquire multi-source frequency domain signals of known fault information of the same or similar mechanical equipment under a first operating condition, and use the multi-source frequency domain signals of the known fault information of the same or similar mechanical equipment under a first operating condition as the source domain dataset; acquire multi-source frequency domain signals of unknown fault information of the mechanical equipment under a second operating condition, and use the multi-source frequency domain signals of the unknown fault information of the mechanical equipment under a second operating condition as the target domain dataset.

[0010] A multi-source dense adaptive adversarial network model is trained using the source domain dataset and the target domain dataset to obtain a mechanical equipment fault diagnosis model. The multi-source dense adaptive adversarial network model includes an unsupervised feature extractor formed by stacking fused convolutional blocks and dense convolutional networks in sequence. The fused convolutional blocks include parallel classical convolutions and dilated convolutions. A fault type classifier and a domain discriminator are set after the unsupervised feature extractor. The loss function of the multi-source dense adaptive adversarial network model is an improved cross-domain adaptive joint loss function.

[0011] The multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition is acquired in real time, and the acquired multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition is input into the mechanical equipment fault diagnosis model, and the fault diagnosis result is output.

[0012] Furthermore, the fault type classifier is formed by stacking a global average pooling layer, a fully connected layer, and a softmax layer in sequence, and the domain discriminator is set after the fully connected layer of the fault type classifier;

[0013] The process of training a multi-source dense adaptive adversarial network model using the source domain dataset and the target domain dataset to obtain a mechanical equipment fault diagnosis model includes:

[0014] The training iteration number of the multi-source dense adaptive adversarial network model is set. The source domain dataset and the target domain dataset are input into the multi-source dense adaptive adversarial network model. The classical convolution and the dilated convolution are both performed on the source domain dataset and the target domain dataset to obtain the source domain classical convolution feature set, the target domain classical convolution feature set, the source domain dilated convolution feature set, and the target domain dilated convolution feature set. The features in the source domain classical convolution feature set and the source domain dilated convolution feature set are concatenated sequentially along the channel dimension to obtain the source domain fusion convolution feature set. The features in the target domain classical convolution feature set and the target domain dilated convolution feature set are concatenated sequentially along the channel dimension to obtain the target domain fusion convolution feature set. The dense convolutional network extracts the deep features in the source domain fusion convolution feature set and the target domain fusion convolution feature set to obtain the source domain deep feature set and the target domain deep feature set.

[0015] The global average pooling layer in the fault type classifier performs global average pooling on the source domain deep feature set and the target domain deep feature set to obtain a source domain global average pooled feature set and a target domain global average pooled feature set. The fully connected layer in the fault type classifier processes the source domain global average pooled feature set and the target domain global average pooled feature set to obtain a source domain fully connected layer feature set and a target domain fully connected layer feature set. The softmax layer in the fault type classifier processes the source domain fully connected layer feature set and the target domain fully connected layer feature set to obtain a source domain fault type dataset and a target domain fault type dataset. The source domain fully connected layer feature set and the source domain fault type dataset are combined through a multilinear mapping operation to form a source domain joint optimized feature set, and the target domain fully connected layer feature set and the target domain fault type dataset are combined through a multilinear mapping operation to form a target domain joint optimized feature set.

[0016] The domain discriminator performs domain discrimination on the source domain joint optimization feature set and the target domain joint optimization feature set to obtain the source domain domain discrimination dataset and the target domain domain discrimination dataset.

[0017] The improved cross-domain adaptive joint loss function is calculated using the source domain fully connected layer feature set, the target domain fully connected layer feature set, the source domain fault type dataset, the source domain sample labels, the target domain fault type dataset, the source domain discriminant dataset, and the target domain discriminant dataset. The parameters of the multi-source dense adaptive adversarial network model are then optimized using the improved cross-domain adaptive joint loss function, the backpropagation algorithm, and the Adam optimizer to complete one training iteration. After training for the specified number of iterations, the mechanical equipment fault diagnosis model is generated. The source domain samples are multi-source frequency domain signals under the first operating condition, and the sample labels of the source domain samples are known fault information corresponding to the multi-source frequency domain signals under the first operating condition.

[0018] Furthermore, the improved cross-domain adaptive joint loss function is:

[0019]

[0020] in,

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] w k = 1 + exp[-H(k)]

[0027]

[0028]

[0029]

[0030]

[0031] Where: δ c δ f and δ g They represent and The parameters of the function to be optimized; α and β are adjustable hyperparameters to balance each loss.

[0032] It is the cross-entropy loss function between the classification prediction probabilities in the source domain fault type dataset and the sample labels of the source domain samples; n s This represents the batch size of samples input from the source domain dataset into the multi-source dense adaptive adversarial network model; m is the number of health states contained in the source domain dataset; 1{y i =j} means when y i The value is 1 when j = 0, otherwise the value is 0; y i p represents the sample label of the i-th source domain sample input from the source domain dataset into the multi-source dense adaptive adversarial network model; j It is the classification prediction probability that the i-th source domain sample in the source domain fault type dataset belongs to the j-th health state;

[0033] d is the depth correlation alignment loss function between the source domain global average pooling feature set and the target domain global average pooling feature set; d is the dimension of the features in the source domain global average pooling feature set and the target domain global average pooling feature set; F M is the Frobenius norm of the square matrix; s and M t n represents the covariance matrices of the source domain global average pooling feature set and the target domain global average pooling feature set, respectively; t is the sample batch size input from the target domain dataset to the multi-source dense adaptive adversarial network model; 1 is a column vector containing only 1 elements; f Ps and f Pt These are deep features derived from the global average pooling feature set of the source domain and the global average pooling feature set of the target domain, respectively.

[0034] It is the entropy-conditional multilinear mapping domain adversarial loss function between the source and target domains, where This represents minimizing the fault type classifier G and the unsupervised feature extractor F. The average value is calculated after processing the source domain samples and their sample labels from the source domain dataset. To calculate the average value of the source domain samples from the source domain dataset after processing; To calculate the average value of target domain samples from the target domain dataset, the target domain samples are multi-source frequency domain signals under the second operating condition; C(·,·) is the cross-entropy loss function; and Let represent the classification prediction probability of the i-th source domain sample in the source domain fault type dataset and the sample label of the source domain sample, respectively; λ is an adjustable hyperparameter that controls the balance between the fault type classifier loss and the domain discriminator loss. This represents the entropy weight of the i-th source domain sample from the source domain dataset. w represents the entropy weight of the j-th target domain sample from the target domain dataset; k Let H(k) be the entropy weights of the source and target domain samples; H(k) represents the entropy of the source and target domain samples; p k is the classification prediction probability that a sample in the source domain fault type dataset and the target domain fault type dataset belongs to the k-th health state; D(·) represents the domain discrimination operation of the domain discriminator; This represents the multilinear mapping result of the u-th source domain sample from the source domain dataset. T(h) represents the multilinear mapping result of the j-th target domain sample from the target domain dataset; T(h) represents the multilinear mapping result. T represents a general multilinear mapping. ⊙ (·,·) denotes a random multilinear mapping; f represents the deep features in the source domain fully connected layer feature set and the target domain fully connected layer feature set; p represents the classification prediction probability in the source domain fault type dataset and the target domain fault type dataset. Represents the multilinear mapping operation between the deep feature f and the classification prediction probability p; ⊙ represents the element-wise product; R f and R p d represents the random matrices sampled from the deep feature f and the classification prediction probability p, respectively; d is the dimension of the multilinear mapping; d f d is the dimension of the deep feature f; p Let p be the dimension of the classification prediction probability.

[0035] Furthermore, the multi-source frequency domain signal used to acquire known fault information of the same or similar mechanical equipment under the first operating condition is:

[0036] Acquire multi-source time-domain signals containing known fault information of mechanical equipment that is the same or similar to the mechanical equipment to be diagnosed under the first working condition;

[0037] The multi-source time-domain signal of the known fault information of the same or similar mechanical equipment under the first working condition is converted by Fast Fourier Transform to obtain the multi-source frequency-domain signal of the known fault information of the same or similar mechanical equipment under the first working condition.

[0038] Furthermore, the multi-source frequency domain signal used to acquire unknown fault information of the mechanical equipment under the second operating condition includes:

[0039] Multi-source time-domain signals for acquiring unknown fault information of the mechanical equipment to be diagnosed under the second operating condition;

[0040] The multi-source time-domain signal of the unknown fault information of the mechanical equipment under the second operating condition is converted by Fast Fourier Transform to obtain the multi-source frequency-domain signal of the unknown fault information of the mechanical equipment under the second operating condition.

[0041] Furthermore, the real-time acquisition of the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second operating condition includes:

[0042] Real-time acquisition of multi-source time-domain signals of the mechanical equipment to be diagnosed under the second operating condition;

[0043] The multi-source time-domain signal of the mechanical equipment under the second operating condition is converted by using Fast Fourier Transform to obtain the multi-source frequency-domain signal of the mechanical equipment under the second operating condition.

[0044] A mechanical equipment fault diagnosis device, comprising:

[0045] The acquisition module is used to acquire multi-source frequency domain signals of known fault information of the same or similar mechanical equipment under a first operating condition, and to use the multi-source frequency domain signals of the known fault information of the same or similar mechanical equipment under the first operating condition as the source domain dataset; and to acquire multi-source frequency domain signals of unknown fault information of the mechanical equipment under a second operating condition, and to use the multi-source frequency domain signals of the unknown fault information of the mechanical equipment under the second operating condition as the target domain dataset.

[0046] The training module is used to train a multi-source dense adaptive adversarial network model using the source domain dataset and the target domain dataset to obtain a mechanical equipment fault diagnosis model. The multi-source dense adaptive adversarial network model includes an unsupervised feature extractor formed by stacking fused convolutional blocks and dense convolutional networks in sequence. The fused convolutional blocks include parallel classical convolutions and dilated convolutions. A fault type classifier and a domain discriminator are set after the unsupervised feature extractor. The loss function of the multi-source dense adaptive adversarial network model is an improved cross-domain adaptive joint loss function.

[0047] The output module is used to acquire the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition in real time, input the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition in real time into the mechanical equipment fault diagnosis model, and output the fault diagnosis result.

[0048] An apparatus includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the mechanical equipment fault diagnosis method.

[0049] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the mechanical equipment fault diagnosis method.

[0050] Compared with the prior art, the present invention has at least the following beneficial effects:

[0051] This invention proposes a fault diagnosis method for mechanical equipment. It uses multi-source frequency domain signals as input to a multi-source dense adaptive adversarial network (DAN) model. Through an advanced network structure design including an unsupervised feature extractor, a fault type classifier, and a domain discriminator, the DAN model achieves stronger feature extraction capabilities with a smaller number of training parameters. It can more effectively integrate and utilize information from multi-source frequency domain signals. A complete domain adaptation strategy, constructed through an improved cross-domain adaptive joint loss function, achieves joint cross-domain optimization of feature and label information. This effectively handles the inherent label information in fault classification problems and more effectively transfers abnormal samples, thus improving the cross-domain intelligent fault diagnosis performance of actual industrial equipment under different operating conditions. This invention constructs an unsupervised feature extractor comprising a dense convolutional network and fused convolutional blocks for extracting deep fusion features from multi-source frequency domain signals. The parallel configuration of classical convolutions and dilated convolutions in the fused convolutional blocks extracts multi-scale deep features from multi-source frequency domain signals and effectively fuses information. Furthermore, the dense connections of the dense convolutional network further fuse multi-source information, enabling the multi-source dense adaptive adversarial network model to achieve deeper modeling with fewer parameters, extracting higher-level features and improving network efficiency. The constructed novel unsupervised feature extractor constitutes a composite fusion strategy at both the data and feature levels, realizing the fusion and effective utilization of multi-source information. The improved cross-domain adaptive joint loss function proposed in this invention employs a deep correlation alignment loss function and an entropy-conditional adversarial domain adaptive strategy. By introducing deep correlation alignment into the network to measure the algorithm, the distribution differences between source domain multi-source data and target domain multi-source data can be effectively reduced. In the entropy-conditional adversarial domain adaptation strategy, the joint optimization of feature and label information in the fault classification task is achieved through the entropy-conditional multilinear mapping domain adversarial loss function. This fully utilizes the inherent label information in the fault classification problem. The adversarial domain adaptation method achieves domain adaptation of the global feature distribution. Applying entropy weights can effectively reduce the impact of training samples with negative transfer effects on domain adaptation. The constructed joint loss function can learn and effectively adapt to multi-source data in both the source and target domains at both the global and local feature distribution levels. This invention, based on a multi-source dense adaptive adversarial network model for mechanical equipment fault diagnosis, combines multi-sensor data-level and feature-level information fusion, automatic learning and extraction of deep features, and adaptive feature domain adaptation methods. It can effectively process information from multiple sensors under different operating conditions of equipment in industrial production, processing and fusing multi-source information in the era of big data in machinery. This provides an effective tool for health monitoring and fault diagnosis of mechanical equipment under varying operating conditions, and is suitable for complex industrial scenarios involving mechanical system fault diagnosis.

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the specific embodiments of the present invention, the drawings used in the description of the specific embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0054] Figure 1 This is a flowchart of a mechanical equipment fault diagnosis method according to the present invention;

[0055] Figure 2 This is a technical flowchart illustrating a specific embodiment of the present invention;

[0056] Figure 3 This is the specific structure of the multi-source dense adaptive adversarial network model of the present invention;

[0057] Figure 4 This is the structure of the fused convolution with dilation rate a = 2 and kernel size d = 2 as described in the specific embodiment of the present invention;

[0058] Figure 5 This refers to the dense block network structure described in a specific embodiment of the present invention;

[0059] Figure 6 The classification results of the CWRU dataset in the specific embodiments of the present invention;

[0060] Figure 7 For the feature visualization of the CWRU dataset in a specific embodiment of the present invention, wherein, (a) CNN,

[0061] (b)DDC, (c)D-CORAL, (d)ADDA, (e)SDAAN, (f)MDAAN;

[0062] Figure 8 The confusion matrix analysis of the CWRU dataset in a specific embodiment of the present invention is shown, where (a) CNN, (b) DDC, (c) D-CORAL, (d) ADDA, (e) SDAAN, and (f) MDAAN. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] As a specific embodiment of the present invention, such as Figure 1 As shown, a method for diagnosing mechanical equipment faults specifically includes the following steps:

[0065] Step 1: Obtain multi-source frequency domain signals of known fault information of the same or similar mechanical equipment under the first operating condition, and use the multi-source frequency domain signals of the known fault information of the same or similar mechanical equipment under the first operating condition as the source domain dataset; Obtain multi-source frequency domain signals of unknown fault information of the mechanical equipment under the second operating condition, and use the multi-source frequency domain signals of the unknown fault information of the mechanical equipment under the second operating condition as the target domain dataset.

[0066] Specifically, the multi-source frequency domain signal containing known fault information of the same or similar mechanical equipment as the one to be diagnosed under the first operating condition is as follows:

[0067] Acquire multi-source time-domain signals containing known fault information of mechanical equipment that is the same or similar to the mechanical equipment to be diagnosed under the first working condition;

[0068] The multi-source time-domain signal of the known fault information of the same or similar mechanical equipment under the first working condition is converted by Fast Fourier Transform to obtain the multi-source frequency-domain signal of the known fault information of the same or similar mechanical equipment under the first working condition.

[0069] Specifically, the multi-source frequency domain signal used to acquire unknown fault information of the mechanical equipment under the second operating condition is as follows:

[0070] Multi-source time-domain signals for acquiring unknown fault information of the mechanical equipment to be diagnosed under the second operating condition;

[0071] The multi-source time-domain signal of the unknown fault information of the mechanical equipment under the second operating condition is converted by Fast Fourier Transform to obtain the multi-source frequency-domain signal of the unknown fault information of the mechanical equipment under the second operating condition.

[0072] Step 2: Train a multi-source dense adaptive adversarial network model using the source domain dataset and the target domain dataset to obtain a mechanical equipment fault diagnosis model. The multi-source dense adaptive adversarial network model includes an unsupervised feature extractor formed by stacking fused convolutional blocks and dense convolutional networks. The fused convolutional blocks include parallel classical convolutions and dilated convolutions. After the unsupervised feature extractor, a fault type classifier is set up, formed by stacking global average pooling layers, fully connected layers, and softmax layers. After the fully connected layer of the unsupervised feature extractor, a domain discriminator is set up. The loss function of the multi-source dense adaptive adversarial network model is an improved cross-domain adaptive joint loss function.

[0073] Specifically, the step of training a multi-source dense adaptive adversarial network model using the source domain dataset and the target domain dataset to obtain a mechanical equipment fault diagnosis model includes:

[0074] a. Set the number of training iterations for the multi-source dense adaptive adversarial network model. Input the source domain dataset and the target domain dataset into the multi-source dense adaptive adversarial network model. Both the classical convolution and the dilated convolution operate on the source domain dataset and the target domain dataset to obtain the source domain classical convolution feature set, the target domain classical convolution feature set, the source domain dilated convolution feature set, and the target domain dilated convolution feature set. Concatenate the features in the source domain classical convolution feature set and the source domain dilated convolution feature set sequentially along the channel dimension to obtain the source domain fusion convolution feature set. Concatenate the features in the target domain classical convolution feature set and the target domain dilated convolution feature set sequentially along the channel dimension to obtain the target domain fusion convolution feature set. The dense convolutional network extracts the deep features in the source domain fusion convolution feature set and the target domain fusion convolution feature set to obtain the source domain deep feature set and the target domain deep feature set.

[0075] b. The global average pooling layer in the fault type classifier performs global average pooling on the source domain deep feature set and the target domain deep feature set to obtain a source domain global average pooled feature set and a target domain global average pooled feature set; the fully connected layer in the fault type classifier processes the source domain global average pooled feature set and the target domain global average pooled feature set to obtain a source domain fully connected layer feature set and a target domain fully connected layer feature set; the softmax layer in the fault type classifier processes the source domain fully connected layer feature set and the target domain fully connected layer feature set to obtain a source domain fault type dataset and a target domain fault type dataset; the source domain fully connected layer feature set and the source domain fault type dataset are combined through a multilinear mapping operation to form a source domain joint optimized feature set, and the target domain fully connected layer feature set and the target domain fault type dataset are combined through a multilinear mapping operation to form a target domain joint optimized feature set;

[0076] c. The domain discriminator performs domain discrimination on the source domain joint optimization feature set and the target domain joint optimization feature set to obtain the source domain domain discrimination dataset and the target domain domain discrimination dataset;

[0077] d. Calculate the improved cross-domain adaptive joint loss function using the source domain fully connected layer feature set, the target domain fully connected layer feature set, the source domain fault type dataset, the source domain sample labels, the target domain fault type dataset, the source domain discriminant dataset, and the target domain discriminant dataset. Then, optimize the parameters of the multi-source dense adaptive adversarial network model using the improved cross-domain adaptive joint loss function, the backpropagation algorithm, and the Adam optimizer to complete one training iteration. After training for the specified number of iterations, generate the mechanical equipment fault diagnosis model. Wherein, the source domain samples are multi-source frequency domain signals under the first operating condition, and the sample labels of the source domain samples are known fault information corresponding to the multi-source frequency domain signals under the first operating condition.

[0078] In this invention, the improved cross-domain adaptive joint loss function is specifically as follows:

[0079]

[0080] in,

[0081]

[0082]

[0083]

[0084]

[0085]

[0086] w k = 1 + exp[-H(k)]

[0087]

[0088]

[0089]

[0090]

[0091] Where: δ c δ f and δ g They represent and The parameters of the function to be optimized; α and β are adjustable hyperparameters to balance each loss;

[0092] It is the cross-entropy loss function between the classification prediction probabilities in the source domain fault type dataset and the sample labels of the source domain samples; n s This represents the batch size of samples input from the source domain dataset into the multi-source dense adaptive adversarial network model; m is the number of health states contained in the source domain dataset; 1{y i =j} means when y i The value is 1 when j = 0, otherwise the value is 0; y i p represents the sample label of the i-th source domain sample input from the source domain dataset into the multi-source dense adaptive adversarial network model; j It is the classification prediction probability that the i-th source domain sample in the source domain fault type dataset belongs to the j-th health state;

[0093] d is the depth correlation alignment loss function between the source domain global average pooling feature set and the target domain global average pooling feature set; d is the dimension of the features in the source domain global average pooling feature set and the target domain global average pooling feature set; F M is the Frobenius norm of the square matrix; s and M t n represents the covariance matrices of the source domain global average pooling feature set and the target domain global average pooling feature set, respectively; t is the sample batch size input from the target domain dataset to the multi-source dense adaptive adversarial network model; 1 is a column vector containing only 1 elements; f Ps and f Pt These are deep features derived from the global average pooling feature set of the source domain and the global average pooling feature set of the target domain, respectively.

[0094] It is the entropy-conditional multilinear mapping domain adversarial loss function between the source and target domains, where This represents minimizing the fault type classifier G and the unsupervised feature extractor F. The average value is calculated after processing the source domain samples and their sample labels from the source domain dataset. To calculate the average value of the source domain samples from the source domain dataset after processing; To calculate the average value of target domain samples from the target domain dataset, the target domain samples are multi-source frequency domain signals under the second operating condition; C(·,·) is the cross-entropy loss function; and Let represent the classification prediction probability of the i-th source domain sample in the source domain fault type dataset and the sample label of the source domain sample, respectively; λ is an adjustable hyperparameter that controls the balance between the fault type classifier loss and the domain discriminator loss. This represents the entropy weight of the i-th source domain sample from the source domain dataset. w represents the entropy weight of the j-th target domain sample from the target domain dataset; k Let H(k) be the entropy weights of the source and target domain samples; H(k) represents the entropy of the source and target domain samples; p k is the classification prediction probability that a sample in the source domain fault type dataset and the target domain fault type dataset belongs to the k-th health state; D(·) represents the domain discrimination operation of the domain discriminator; This represents the result of a multilinear mapping from the i-th source domain sample in the source domain dataset. T(h) represents the multilinear mapping result of the j-th target domain sample from the target domain dataset; T(h) represents the multilinear mapping result. T represents a general multilinear mapping. ⊙ (·,·) denotes a random multilinear mapping; f represents the deep features in the source domain fully connected layer feature set and the target domain fully connected layer feature set; p represents the classification prediction probability in the source domain fault type dataset and the target domain fault type dataset. Represents the multilinear mapping operation between the deep feature f and the classification prediction probability p; ⊙ represents the element-wise product; R f and R p d represents the random matrices sampled from the deep feature f and the classification prediction probability p, respectively; d is the dimension of the multilinear mapping; d f d is the dimension of the deep feature f; p Let p be the dimension of the classification prediction probability.

[0095] Step 3: Acquire the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition in real time, input the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition in real time into the mechanical equipment fault diagnosis model, and output the fault diagnosis result.

[0096] Specifically, the real-time acquisition of the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second operating condition includes:

[0097] Real-time acquisition of multi-source time-domain signals of the mechanical equipment to be diagnosed under the second operating condition;

[0098] The multi-source time-domain signal of the mechanical equipment under the second operating condition is converted by using Fast Fourier Transform to obtain the multi-source frequency-domain signal of the mechanical equipment under the second operating condition.

[0099] To provide a more detailed explanation of this invention, the following description uses the Case Western Reserve University bearing dataset as an example to illustrate a complete and detailed method for diagnosing mechanical equipment faults as described in this invention. The technical flow of this specific embodiment is as follows: Figure 2 As shown, the specific description is as follows.

[0100] The bearing tests were conducted under four different operating conditions, including 0 (L0), 1 (L1), 2 (L2), and 3 (L3) horsepower motor loads. Three types of faulty bearings were treated in the tests: rolling element failure (RF), inner ring failure (IR), and outer ring failure (OR). Each fault type consisted of three different sizes: 7 inches, 14 inches, and 21 inches. Including the bearings in normal condition (NO), there were a total of 10 healthy bearing conditions.

[0101] Step 1: This time, a dataset with a sampling frequency of 12kHz is used. Any one of the four working conditions is taken as the first working condition, and any other working condition different from it is taken as the second working condition. A total of twelve cross-domain intelligent fault diagnosis tasks are set up, namely L0→L1, L0→L2, L0→L3, L1→L0, L1→L2, L1→L3, L2→L0, L2→L1, L2→L3, L3→L0, L3→L1, L3→L2.

[0102] Monitoring data from the drive end and wind turbine end under various operating conditions were selected as multi-source time-domain data for the model proposed in this invention. This data includes 10 health states, with 400 samples per health state under each operating condition. 200 samples were allocated to the training set, 100 to the test set, and 100 to the validation set to find the optimal model parameters during iteration. Each sample contains 1024 sampling points. All comparison methods used the single-source time-domain signal from the drive end as input. Specifically, the multi-source time-domain signals of known fault information in the training and validation sets under the first operating condition in each cross-domain intelligent fault diagnosis task are the multi-source time-domain signals of known fault information of the mechanical equipment under the first operating condition. The multi-source time-domain signals of unknown fault information in the training set under the second operating condition are the multi-source time-domain signals of unknown fault information of the mechanical equipment under the second operating condition. The multi-source time-domain signals of known fault information in the validation set under the first operating condition in each cross-domain intelligent fault diagnosis task are used as the validation set during model training to verify the model's state and search for the optimal hyperparameters.

[0103] A Fast Fourier Transform (FFT) is used to transform the multi-source time-domain signal of the mechanical equipment under the first operating condition, resulting in a multi-source frequency-domain signal of the known fault information under the first operating condition, which serves as the source domain dataset. Similarly, a Fast Fourier Transform (FFT) is used to transform the multi-source time-domain signal of the mechanical equipment under the second operating condition, resulting in a multi-source frequency-domain signal of the unknown fault information under the second operating condition, which serves as the target domain dataset.

[0104] Step 2: Construct a Multi-Source Dense Adaptive Adversarial Network (MDAAN) model and train it using the source domain dataset and the target domain dataset. The detailed structure of the constructed MDAAN model is as follows: Figure 3 As shown in Table 1, the network structure of the entire model is configured as follows.

[0105] Table 1. Network Structure Settings for the MDAAN Model

[0106]

[0107] The specific structural descriptions of important subnetworks in the MDAAN model network are as follows:

[0108] 1) The network constructs an unsupervised feature extractor. The unsupervised feature extractor combines fused convolutional and dense convolutional networks, including a fused convolutional layer, three dense-transition blocks, and a remaining dense block, which can extract deeper features of the signal while expanding the receptive field.

[0109] 1.1) The specific implementation process of fusion convolution is as follows: Figure 4 As shown, fusion convolution involves two types of convolution arranged in parallel: standard convolution and atrous convolution. Atrous convolution adds a dilated rate hyperparameter 'a' to the classic convolution, allowing it to obtain multi-scale information of the input features without adding extra parameters. Due to its ability to arbitrarily expand the receptive field, atrous convolution works well for input images that require more global information. Unlike classic convolution operations, atrous convolution determines the size of the discontinuous convolution kernel based on the dilation rate 'a'. Assuming 'q' is the size of the classic convolution kernel, the formula for calculating the actual kernel size of atrous convolution is as follows.

[0110] Q = q + (q - 1)(a - 1) (1)

[0111] Where Q represents the actual kernel size of the dilated convolution.

[0112] However, improper dilation rate adjustment or the use of only dilated convolutions in the model can lead to the loss of some original information. Therefore, classical convolutions and dilated convolutions are arranged in parallel and concatenated along the channel dimension to form a fused convolution to overcome the above shortcomings.

[0113] 1.2) The dense convolutional network mainly consists of two parts: dense blocks and transition blocks. In the model constructed in this invention, the growth rate k and the compression factor θ are set to 25 and 0.5, respectively.

[0114] 1.2.1) Dense Blocks: The structure of dense blocks is as follows: Figure 5 As shown, the dense layers of a dense block consist of stacked batch normalization (BN) layers, ReLU layers, and convolutional layers. The first dense layer of each dense block receives the input feature map χ0 as input, which is then sequentially fed into the BN layer, ReLU layer, and convolutional layer to extract a new feature map χ1. Then, χ0 and χ1 are concatenated as the input to the second dense layer, and the same steps are performed to obtain a new feature map χ2. Afterward, the existing feature maps are concatenated after each dense layer. The specific formula can be expressed as follows.

[0115] Ii = concat([χ0,…,χ) i (2)

[0116] Where χ0,…,χ i These are the feature maps obtained from the i-th convolution, and concat(·) represents the concatenation operation. i Represents the connection feature map [χ0,…,χ] based on the i-th connection operation. i The result.

[0117] Simultaneously, a growth rate factor k is defined within the dense block to enrich different types of internal features. Assuming the number of channels in the i-th convolutional layer of the block is k, the number of feature maps input to the i-th dense layer is:

[0118] L i =L0+k(i-1) (3)

[0119] Where L i L0 represents the number of feature maps input to the i-th dense layer, and L0 represents the number of original feature maps input to the first dense layer.

[0120] 1.2.2) Transition Block: The structure of the transition block is similar to the pooling layer in a classic CNN, both aimed at reducing the size of the feature map and the number of parameters. A transition block can consist of a Batch Normalization (BN) layer, a ReLU layer, a convolutional layer, and an average pooling layer. Furthermore, to further reduce the number of parameters, a compression factor θ is defined in the transition block.

[0121] 2) The network constructs a fault type classifier after the unsupervised feature extractor, consisting of a globally average pooling layer, a fully connected layer, and a softmax layer stacked sequentially. For health state classification and two-layer domain adaptation, a globally average pooling layer and a fully connected layer are applied after the feature extractor. At the end of the model, a softmax layer serves as the network's output. The number of neurons in this layer corresponds to the number of healthy states present in the input data, enabling health state classification.

[0122] 3) A domain discriminator D is constructed after the fully connected layer of the fault type classifier. The parameters and structure of the domain discriminator D are described in reference "M. Long, Z. Cao, J. Wang, and MI Jordan, 'Conditional adversarial domain adaptation,' in Adv. neural inf.proces.syst., 2018, pp. 1640-1650." The discriminator D consists of three fully connected layers with 1024, 1024, and 1 neuron, respectively.

[0123] After constructing the Multi-Source Dense Adaptive Adversarial Network (MDAAN) model, the constructed MDAAN model was trained using the training data. The batch size was set to 500, and the number of training iterations for all datasets was set to 150. During training, an improved cross-domain adaptive joint loss function, backpropagation algorithm, and learning rate of 1×10⁻⁶ were used. -3 The Adam optimizer optimizes the model's parameters. All methods use a validation set to search for the optimal hyperparameters. After training for a set number of iterations, a mechanical equipment fault diagnosis model is generated. More specifically, in the improved cross-domain adaptive joint loss function, α and β are adjustable hyperparameters used to balance each loss. This paper selects a value of 1 for β and chooses the value of α using the following mathematical formula:

[0124]

[0125] Where q is a factor that changes linearly from 0 to 1 during the training process.

[0126] When training the model using training data, the training optimization objective is adjusted. Through continuous optimization, a high-performance mechanical equipment fault diagnosis model, trained for various operating conditions, is finally generated, achieving cross-domain fault diagnosis. Furthermore, each method is repeated ten times to eliminate the influence of random factors.

[0127] Step 3: Use Fast Fourier Transform to convert the multi-source time-domain signals of the known fault information in the test set under the second working condition to obtain the multi-source frequency-domain signals of the test set under the second working condition, which will serve as the model test set. Input the model test set into the trained mechanical equipment fault diagnosis model to obtain the fault diagnosis results, and further evaluate the model's performance.

[0128] To verify the superiority of the mechanical equipment fault diagnosis model provided by this invention, this embodiment selects four high-level domain adaptive methods based on deep networks for comparison. These four methods include:

[0129] 1. Deep Convolutional Neural Networks (CNNs), as the most common pattern recognition method, have a network structure shown in Table 2. The entire network employs Batch Normalization (BN) layers and dropout techniques to enhance network performance, using a learning rate of 1×10⁻⁶. -3 The Adam optimizer optimizes the model.

[0130] Table 2. CNN Network Structure

[0131]

[0132] 2. Deep Domain Confusion (DCC), detailed process reference "E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Japa Darrell, 'Deep domain confusion: Maximizing for domain invariance,' 2014." This method applies MMD to deep networks to achieve domain adaptation between the source and target domains. A learning rate of 2×10⁻⁶ is used. -2 A mini-batch stochastic gradient descent (SGD) optimizer with momentum of 0.9 is used to optimize the entire DDC model.

[0133] 3. Deep Correlation Alignment (D-CORAL) for Deep Domain Adaptation: For detailed procedures, refer to "B. Sun and K. Saenko, 'Deep coral: Correlation alignment for deep domain adaptation,' in Lect. NotesComput.Sci., 2016, pp.443-450: Springer." Deep CORAL extends the CORAL algorithm, utilizing more powerful nonlinear transformations and applying it to deep networks. A learning rate of 1×10⁻⁶ is used. -4 The Adam optimizer optimizes the entire network.

[0134] 4. Adversarial Discriminative Domain Adaptation (ADDA). For detailed procedures, refer to "E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell, 'Adversarial discriminative domain adaptation,' in Proc.-IEEE Conf. Comput. Vis. Pattern Recognit, CVPR, 2017, pp. 7167-7176." ADDA, as an advanced adversarial domain adaptation method, makes adversarial transfer more suitable for classification problems with large domain distribution shifts. In this training process, a learning rate of 1×10⁻⁶ was used. -3 The Adam optimizer optimizes the model.

[0135] In addition, a Single-Source Dense Adaptive Adversarial Network (SDAAN) was used as a comparison method. Its network structure is the same as MDAAN, but it uses a single-source signal as input. To compare with multi-source signal input, the fused convolutional layer was replaced with a regular convolutional layer. A learning rate of 1×10⁻⁶ was chosen. -3 The Adam optimizer is used to optimize the model. The comparison results of the six models are as follows:

[0136] 1) Classification accuracy and experimental results are as follows Figure 6 As shown in Table 3, it is evident that the MDAAN model and SDANN proposed in this invention achieve the highest average accuracy, reaching 99.73% and 99.39%, respectively. In contrast, among the other four comparison methods, the highest average accuracy is 94.93% for ADDA, approximately 5% lower than the proposed method. However, due to the relatively simple signal composition of this dataset, the accuracy of MDAAN is not significantly higher than that of SDAAN. Furthermore, MDAAN and SDAAN achieve the best results among all methods when optimizing relatively few parameters. This demonstrates the superiority of the proposed method.

[0137] Table 3. Classification results (%) of the CWRU dataset

[0138]

[0139] 2) This invention randomly selects one test result from ten L3→L1 task tests to analyze the domain adaptability of each method. The test result is then visualized using t-SNE, and the feature distribution is as follows: Figure 7 As shown, from Figure 7 As shown in (e)-(f), the MDAAN and SDAAN models proposed in this paper fit the distribution of the 10 health state samples in the source and target domains very well, and the samples of each health state are well separated with clear classification boundaries. In contrast, according to... Figure 7While the other four methods (a)-(d) fit the sample distributions of the source and target domains to some extent, their fitting ability remains limited, leading to many misclassifications. Therefore, these results demonstrate the powerful classification and recognition capabilities and domain adaptability of the method proposed in this invention.

[0140] 3) Based on the same task, further observe the classification results using a confusion matrix. According to... Figure 8 It can be seen that the comparison methods CNN, DDC, D-CORAL, ADDA, and SDAN all exhibit varying degrees of classification error in identifying rolling element fault categories. Furthermore, these methods show significant fluctuations and lag behind MDAAN to some extent. In contrast, MDAAN demonstrates a clear advantage, achieving a considerably high classification accuracy and exhibiting better robustness.

[0141] The present invention also provides a mechanical equipment fault diagnosis device, comprising:

[0142] The acquisition module is used to acquire multi-source frequency domain signals of known fault information of the same or similar mechanical equipment under a first operating condition, and to use the multi-source frequency domain signals of the known fault information of the same or similar mechanical equipment under the first operating condition as the source domain dataset; and to acquire multi-source frequency domain signals of unknown fault information of the mechanical equipment under a second operating condition, and to use the multi-source frequency domain signals of the unknown fault information of the mechanical equipment under the second operating condition as the target domain dataset.

[0143] The training module is used to train a multi-source dense adaptive adversarial network model using the source domain dataset and the target domain dataset to obtain a mechanical equipment fault diagnosis model. The multi-source dense adaptive adversarial network model includes an unsupervised feature extractor formed by stacking fused convolutional blocks and dense convolutional networks in sequence. The fused convolutional blocks include parallel classical convolutions and dilated convolutions. A fault type classifier and a domain discriminator are set after the unsupervised feature extractor. The loss function of the multi-source dense adaptive adversarial network model is an improved cross-domain adaptive joint loss function.

[0144] The output module is used to acquire the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition in real time, input the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition in real time into the mechanical equipment fault diagnosis model, and output the fault diagnosis result.

[0145] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used to implement the operation of a mechanical equipment fault diagnosis method.

[0146] In one embodiment of the present invention, if a mechanical equipment fault diagnosis method is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data.

[0147] The computer storage medium can be any available medium or data storage device that a computer can access, including but not limited to magnetic storage (e.g., floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs)), optical storage (e.g., CDs, DVDs, BDs, HVDs), and semiconductor storage (e.g., ROMs, EPROMs, EEPROMs, non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0148] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0149] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0150] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0151] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0152] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A mechanical equipment failure diagnosis method characterized by, The method comprises the following steps: obtaining multi-source frequency domain signals of known fault information of mechanical equipment of the same type as the mechanical equipment to be diagnosed under a first working condition, taking the multi-source frequency domain signals of known fault information of the mechanical equipment of the same type as the mechanical equipment to be diagnosed under the first working condition as a source domain data set; obtaining multi-source frequency domain signals of unknown fault information of the mechanical equipment to be diagnosed under a second working condition, taking the multi-source frequency domain signals of unknown fault information of the mechanical equipment to be diagnosed under the second working condition as a target domain data set; training a multi-source dense adaptive adversarial network model by using the source domain data set and the target domain data set to obtain a mechanical equipment fault diagnosis model, wherein the multi-source dense adaptive adversarial network model comprises an unsupervised feature extractor formed by sequentially stacking a fusion convolution block and a dense convolution network, the fusion convolution block comprises a classical convolution and a hollow convolution arranged in parallel, the unsupervised feature extractor is provided with a fault type classifier and a domain discriminator, the fault type classifier is formed by sequentially stacking a global average pooling layer, a fully connected layer and a softmax layer, and the domain discriminator is arranged after the fully connected layer of the fault type classifier; the training process comprises the following steps: setting the number of training iterations of the multi-source dense adaptive adversarial network model, inputting the source domain data set and the target domain data set into the multi-source dense adaptive adversarial network model, performing convolution operation on the source domain data set and the target domain data set by using the classical convolution and the hollow convolution to obtain a source domain classical convolution feature set, a target domain classical convolution feature set, a source domain hollow convolution feature set and a target domain hollow convolution feature set, sequentially concatenating features in the source domain classical convolution feature set and the source domain hollow convolution feature set in the channel dimension to obtain a source domain fusion convolution feature set, sequentially concatenating features in the target domain classical convolution feature set and the target domain hollow convolution feature set in the channel dimension to obtain a target domain fusion convolution feature set, and extracting deep features in the source domain fusion convolution feature set and the target domain fusion convolution feature set by using the dense convolution network to obtain a source domain deep feature set and a target domain deep feature set; the loss function of the multi-source dense adaptive adversarial network model is an improved cross-domain adaptive joint loss function, and the improved cross-domain adaptive joint loss function is as follows: wherein, , and respectively represent , and parameters to be optimized for the function; and are adjustable hyperparameters; is a cross-entropy loss function between the classification prediction probability of the source domain fault type dataset and the sample label of the source domain sample; is a deep correlation alignment loss function between the source domain global average pooling feature set and the target domain global average pooling feature set; is an entropy conditional multi-linear mapping domain adversarial loss function between the source domain and the target domain, wherein represents minimizing the fault type classifier and the unsupervised feature extractor , is an average value obtained by processing the source domain sample from the source domain dataset and the sample label of the source domain sample; is an average value obtained by processing the source domain sample from the source domain dataset; is an average value obtained by processing the target domain sample from the target domain dataset, the target domain sample being a multi-source frequency domain signal under a second working condition; is a cross-entropy loss function; respectively represent the classification prediction probability of the i th source domain sample in the source domain fault type dataset and the sample label of the source domain sample; is an adjustable hyperparameter that controls the balance between the fault type classifier loss and the domain discriminator loss; represents the entropy weight of the th source domain sample from the source domain dataset, represents the entropy weight of the th target domain sample from the target domain dataset; is the entropy weight of the source domain sample and the target domain sample; represents the entropy of the source domain sample and the target domain sample; is the classification prediction probability that the samples in the source domain fault type dataset and the target domain fault type dataset belong to the th health state; represents the domain discrimination operation of the domain discriminator; represents the multi-linear mapping result of the th source domain sample from the source domain dataset, represents the multi-linear mapping result of the a multi-linear mapping result of a target domain sample; denotes a multi-linear mapping result; denotes a common multi-linear mapping, denotes a random multi-linear mapping; is a deep feature in a source domain fully connected layer feature set and a target domain fully connected layer feature set; is a classification prediction probability in the source domain fault type data set and the target domain fault type data set; denotes a deep feature and a classification prediction probability ; denotes an element-wise product; and represent a random matrix sampled from a deep feature and a classification prediction probability ; d is a dimension of a multi-linear mapping; is a dimension of a deep feature ; is a dimension of a classification prediction probability ; obtaining multi-source frequency domain signals of the mechanical equipment to be diagnosed under the second working condition in real time, inputting the multi-source frequency domain signals of the mechanical equipment to be diagnosed under the second working condition obtained in real time into the mechanical equipment fault diagnosis model, and outputting a fault diagnosis result.

2. The method of claim 1, wherein, The global average pooling layer in the fault type classifier performs global average pooling processing on the source domain deep feature set and the target domain deep feature set to obtain a source domain global average pooling feature set and a target domain global average pooling feature set; the fully connected layer in the fault type classifier processes the source domain global average pooling feature set and the target domain global average pooling feature set to obtain a source domain fully connected layer feature set and a target domain fully connected layer feature set; the softmax layer in the fault type classifier processes the source domain fully connected layer feature set and the target domain fully connected layer feature set to obtain a source domain fault type data set and a target domain fault type data set; the source domain fully connected layer feature set and the source domain fault type data set are combined into a source domain joint optimization feature set through a multi-linear mapping operation, and the target domain fully connected layer feature set and the target domain fault type data set are combined into a target domain joint optimization feature set through a multi-linear mapping operation; The domain discriminator performs domain discrimination on the source domain joint optimization feature set and the target domain joint optimization feature set to obtain a source domain domain discrimination data set and a target domain domain discrimination data set; The improved cross-domain adaptive joint loss function is calculated using the source domain fully connected layer feature set, the target domain fully connected layer feature set, the source domain fault type data set, the sample label of the source domain sample, the target domain fault type data set, the source domain domain discrimination data set and the target domain domain discrimination data set, and the parameters of the multi-source dense adaptive adversarial network model are optimized using the improved cross-domain adaptive joint loss function, a back propagation algorithm and an Adam optimizer to complete one training, and the mechanical equipment fault diagnosis model is generated after the training for the training iteration number of times; wherein the source domain sample is a multi-source frequency domain signal under a first working condition, and the sample label of the source domain sample is known fault information corresponding to the multi-source frequency domain signal under the first working condition.

3. The mechanical equipment fault diagnosis method of claim 2, wherein: represents the sample batch size input from the source domain data set to the multi-source dense adaptive adversarial network model; represents the number of health states contained in the source domain data set; represents 1 when , otherwise 0; represents the sample label of the i th source domain sample input from the source domain data set to the multi-source dense adaptive adversarial network model; is the classification prediction probability that the i th source domain sample in the source domain fault type data set belongs to the j th health state; . is the dimension of the features in the source domain global average pooling feature set and the target domain global average pooling feature set; is the Frobenius norm of a square matrix; and respectively represent the covariance matrix of the source domain global average pooling feature set and the target domain global average pooling feature set; is the batch size of samples input to the multi-source dense adaptive adversarial network model from the target domain data set; is a column vector containing all 1 elements; and are deep features from the source domain global average pooling feature set and the target domain global average pooling feature set, respectively.

4. The method of claim 1, wherein, the known fault information of the multi-source frequency domain signal of the same or same type mechanical equipment as the mechanical equipment to be diagnosed under the first working condition is obtained; a known fault information of a multi-source time domain signal of the same or same type mechanical equipment as the mechanical equipment to be diagnosed under the first working condition is obtained; the known fault information of the multi-source time domain signal of the same or same type mechanical equipment as the mechanical equipment to be diagnosed under the first working condition is converted into the known fault information of the multi-source frequency domain signal of the same or same type mechanical equipment as the mechanical equipment to be diagnosed under the first working condition by using fast Fourier transform.

5. The method of claim 1, wherein, the unknown fault information of the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition is obtained, including: a unknown fault information of a multi-source time domain signal of the mechanical equipment to be diagnosed under the second working condition is obtained; the unknown fault information of the multi-source time domain signal of the mechanical equipment to be diagnosed under the second working condition is converted into the unknown fault information of the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition by using fast Fourier transform.

6. The mechanical equipment fault diagnosis method according to claim 1, characterized in that, The real-time acquisition of the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition comprises: Real-time acquisition of multi-source time domain signals of the mechanical equipment to be diagnosed under the second working condition; The multi-source time domain signals of the mechanical equipment to be diagnosed under the second working condition are converted by using fast Fourier transform to obtain the multi-source frequency domain signal of the mechanical equipment to be diagnosed under the second working condition.

7. A mechanical equipment failure diagnosis device characterized by comprising: A mechanical equipment fault diagnosis method for realizing any one of claims 1-6 comprises: An acquisition module is configured to acquire multi-source frequency domain signals of known fault information of the same or same type mechanical equipment as the mechanical equipment to be diagnosed under a first working condition, and use the multi-source frequency domain signals of known fault information of the same or same type mechanical equipment as the mechanical equipment to be diagnosed under the first working condition as a source domain data set; and acquire multi-source frequency domain signals of unknown fault information of the mechanical equipment to be diagnosed under a second working condition, and use the multi-source frequency domain signals of unknown fault information of the mechanical equipment to be diagnosed under the second working condition as a target domain data set; A training module is configured to train a multi-source dense adaptive adversarial network model by using the source domain data set and the target domain data set to obtain a mechanical equipment fault diagnosis model, wherein the multi-source dense adaptive adversarial network model comprises an unsupervised feature extractor formed by sequentially stacking a fusion convolution block and a dense convolution network, the fusion convolution block comprises a classical convolution and a hollow convolution arranged in parallel, the unsupervised feature extractor is provided with a fault type classifier and a domain discriminator, and a loss function of the multi-source dense adaptive adversarial network model is an improved cross-domain adaptive joint loss function; An output module is configured to acquire real-time multi-source frequency domain signals of the mechanical equipment to be diagnosed under the second working condition, input the real-time multi-source frequency domain signals of the mechanical equipment to be diagnosed under the second working condition into the mechanical equipment fault diagnosis model, and output a fault diagnosis result.

8. An apparatus comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, The processor executes the computer program to realize the steps of the mechanical equipment fault diagnosis method according to any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by the processor to realize the steps of the mechanical equipment fault diagnosis method according to any one of claims 1-6.

Images