A method for intelligent diagnosis of device status using distributed multi-source sensing

Abstract

This invention discloses a distributed multi-source sensing intelligent diagnostic method for equipment status. The method includes: 1) Distributed optimization of multiple contact sensors based on the kinematic relationships, dynamic models, and optimization objectives of the mechanical equipment; 2) Acquiring distributed multi-source sensing feedback data; 3) Obtaining standard preprocessed feature data through data fusion; 4) Obtaining distributed multi-source contact equipment status intelligent diagnostic results through transfer learning on the standard feature data; 5) Deploying a mobile dynamic image acquisition and analysis system to implement image-based non-contact equipment status extraction, focusing on the external representation of equipment motion; 6) Extracting key node features through an image analysis system; 7) Obtaining intelligent diagnostic results for the mechanical device status using a graph convolutional neural network; 8) Based on the distributed multi-source contact diagnostic results and the mobile non-contact image diagnostic results, using expert fuzzy fusion technology to obtain the classification and location of the mechanical device's motion status and abnormal conditions, and to predict the equipment's lifespan.

Description

Technical Field

[0001] This invention is applied in the field of intelligent diagnosis of the condition of mechanical equipment, and relates to an intelligent diagnosis method for the condition of equipment using distributed multi-source sensing. Background Technology

[0002] With the rapid development of science and technology and intelligent manufacturing, manufacturing equipment, robots, and other mechanical devices are being applied in an increasing number of production scenarios. To ensure the reliability of these devices, real-time diagnostic results of their operating status are necessary for timely adjustments to strategies and tasks. Placing an appropriate number of sensors at key locations on the equipment and in its surrounding environment can collect information related to its operating status. Proper sensor placement is of great significance.

[0003] Current data acquisition platforms for mechanical equipment typically deploy a limited number of electrical and vibration signal detection modules only for core drive and transmission components, concentrated in key areas. Due to advancements in materials and manufacturing processes, the frequency of malfunctions in purely core functional components is no longer significant; however, the possibility and adverse consequences of malfunctions in other load-bearing and actuating components cannot be ignored. Furthermore, the comprehensive acquisition of the overall operating status of the mechanical equipment is crucial for the effective implementation of production tasks. With the emergence of numerous monitoring points, issues such as optimized sensor placement and multi-source data processing have become apparent. Blindly or intuitously deploying a large number of sensors can lead to excessive redundancy and duplication, hindering the acquisition of comprehensive and high-quality information, and slowing down real-time data processing. Therefore, it is necessary to optimize the distribution and placement of multi-source sensors to obtain effective diagnostic results for the mechanical equipment's condition.

[0004] Currently, acquiring equipment status often involves deploying various sensor modules on the equipment. However, the acquired data is always limited by the information that the equipment itself can transmit and the types of sensors available, making it difficult to assess newly emerging conditions. Obtaining image information of the equipment through visual means can intuitively demonstrate its motion status and abnormal conditions, effectively supplementing the above process. By deploying multiple fixed-position cameras around the equipment and collecting images from a limited number of points, information beyond the dimensions of traditional methods can be obtained. However, visual inspection is only suitable for relatively obvious shape and position differences and foreign object interference, and it is difficult to detect problems such as wear, pitting, and loose parts. In addition, acquiring comprehensive image information requires a large number of fixed-point cameras, posing various challenges in investment and application. Designing a non-fixed visual inspection system and effectively integrating it with a distributed multi-source sensing detection system can fully overcome the aforementioned shortcomings, thereby effectively ensuring the continuous, efficient, and stable operation of mechanical equipment. Summary of the Invention

[0005] The present invention aims to provide a method for intelligent diagnosis of device status using distributed multi-source sensing, in order to overcome the shortcomings of current intelligent diagnosis technologies.

[0006] To address the aforementioned problems and achieve optimized layout of distributed multi-source sensing and intelligent diagnosis of mechanical equipment status, this invention provides a distributed multi-source sensing-based intelligent equipment status diagnosis method. This method will facilitate timely feedback of status and comprehensive control over abnormal situations, effectively preventing accidents caused by untimely handling of mechanical equipment abnormalities and improving the safety and reliability of the entire mechanical equipment.

[0007] To achieve the purpose of the appeal, the present invention adopts the following technical solution:

[0008] Based on the global coordinate system, the kinematic relationship of the mechanical equipment is solved and the dynamic model of the equipment is established. Vibration, acceleration and strain sensors are arranged in the basic units of the equipment such as drive device, actuator, load-bearing component and transmission component. In each electric drive actuator, voltage and current sensing modules are used to establish signal detection channels.

[0009] For each basic unit, establish a rectangular coordinate system or cylindrical coordinate system that adapts to the shape and force characteristics of the unit. Construct an objective function for the distribution of sensors on the surface of the unit. Use the gradient descent method to solve for the optimal position and number of sensors distributed on the surface of the basic unit. For units that cannot be solved, arrange them in a centrally symmetrical manner with an odd number of sensors.

[0010] Based on the device's dynamic model, kinematic relationships, and the layout of contact sensors, a tethered mobile platform is used to deploy a non-contact image acquisition system. This tethered mobile platform includes, but is not limited to, a multi-rotor UAV equipped with a binocular camera and a light source. Multiple image acquisition points are set up to assess the spatial layout of the device and the external representation of the motion of each basic unit, enabling real-time dynamic status monitoring and image capture. The A* algorithm is used to establish the optimal motion path, achieving efficient acquisition by the non-contact image acquisition system.

[0011] The contact signal analysis system uses raw data collected by the contact sensors of the equipment as input, and uses Fast Fourier Transform and Classical Mode Decomposition to extract signal features. It extracts the time-frequency domain features of the signal through signal processing techniques such as denoising, frequency reduction, decomposition, and compression. Through a multi-scale mechanism, the extracted signal features and the raw signal are combined into three-channel data to achieve contact multi-source sensor distributed detection data fusion.

[0012] Fast Fourier Transform (FFT): This method can effectively and quickly transform signals to obtain information about each frequency component in the frequency domain. By removing high-frequency signals from the original signal, a filtered original signal is obtained. Its mathematical formula is:

[0013]

[0014] Where x(n) is a discrete signal of finite length, 0≤k≤N-1.

[0015] Classical mode decomposition (CMD) is commonly used to process nonlinear and non-stationary signals. It decomposes a continuous and complex original signal into multiple eigenmode functions over a time scale to represent frequency components at different time scales. Its mathematical formula is:

[0016]

[0017] Among them, c i r(t) is the finite-length Intrinsic Mode Function (IMF), and r(t) is the signal data residual.

[0018] The raw signals acquired by the contact sensor are processed through Fast Fourier Transform and Classical Mode Decomposition to form a three-channel signal data set, which is then combined with the original data. A length of N is extracted from each channel data set. 2 The continuous signal is converted into N×N two-dimensional image data and then processed to grayscale. Weight parameters a1, a2, and a3 are then assigned to the three channels, where a1, a2, and a3 satisfy the following relationship:

[0019] a1 + a2 + a3 = 1, 0 ≤ a1, a2, a3 ≤ 1

[0020] The grayscale images corresponding to the three channels are then superimposed to obtain a three-channel two-dimensional data image, and the data is fused through multi-scale weighting.

[0021] Because environmental noise interferes with signal data, the data distribution shift caused by abnormal data is much smaller than that caused by normal data during acquisition, making it difficult to collect effective labeled data from sensors or to effectively label the data. Consequently, state diagnosis based on deep learning and traditional signal processing methods is usually ineffective. To address these issues, this invention employs a deep convolutional adversarial domain adaptive neural network model (DCADA) to achieve intelligent equipment state diagnosis under unsupervised, noisy, and data distribution shift conditions.

[0022] The raw signal data acquired by the contact sensor is preprocessed through data fusion, resulting in two domain sample data sets: a source domain and a target domain. The source domain samples contain labeled signal feature data, while the target domain samples contain unlabeled signal feature data. These segmented domain sample data are then used as input to the DCADA model.

[0023] The DCADA model comprises feature extraction, domain adaptation, and state recognition modules. The feature extraction module includes five convolutional modules, each consisting of a convolutional layer, an activation function, and a pooling layer. The feature extraction module is finally connected to a fully connected layer. The domain adaptation module includes multiple domain discriminators and a marginal distribution difference measure function (MMD). The state recognition module includes a convolutional block, a fully connected layer, and a normalized exponential softmax function.

[0024] The feature extraction module converts domain sample signals into learnable classification information. The loss function formula for its feature extraction module is:

[0025]

[0026] The domain adaptation module reduces the distributional differences of domain samples and extracts domain-invariant features. It employs multiple domain discriminators to classify source and target domain information. The inter-class distribution distance is calculated using Maximum Mean Difference (MMD), and the resulting loss value is used as an adaptation factor to guide domain adaptation for marginal distributional differences. The domain discriminator, as a standard binary classification network, includes global average pooling, convolutional blocks, and a softmax function module. It obtains the probability values ​​corresponding to the feature information as the output of the domain discriminator, and its loss function is:

[0027]

[0028] Where θ f For G f The weights are iteratively updated to update the parameters.

[0029] Maximum Mean Difference (MMD) embeds high-level features from two domains into two points in the Reproducing Kernel Hilbert Space (RKHS) by introducing a mapping function. The distance between these two points represents the distributional distance between the features of the two domains. Matching the distributions of the two domains is achieved by minimizing this distance, where the distributional distance can be defined as...

[0030]

[0031] Where Φ is the mapping function of RKHS, X s With X t Let L represent samples from the source and target domains, respectively, and let L be the loss function of MMD. M :

[0032]

[0033] Where k(x,y) is a high-dimensional mapping space.

[0034] State recognition module: This module classifies features extracted adaptively from the domain to achieve anomaly classification and diagnosis. It includes convolutional blocks, fully connected layers, and a normalized softmax function. Its loss function is L... y for:

[0035]

[0036] The optimization objective for DCADA model parameter calculation is:

[0037]

[0038]

[0039] in Representing θ f ,θ c ,θ d The parameter estimates, θ f ,θ d ,θ c For G f G d G y The weights are iteratively updated to update the parameters.

[0040] The iterative update formula for DCADA model parameters is:

[0041]

[0042]

[0043]

[0044] Where μ is the learning rate hyperparameter.

[0045] Non-contact image acquisition systems take non-contact acquired image data as input, extract image features of a given acquisition point through image processing techniques such as segmentation, denoising, enhancement, and grayscale, and label the external representations of the extracted basic units.

[0046] Due to the nonlinear and irregular nature of image data, deep learning operations such as convolution and pooling cannot be directly applied to image data. Furthermore, since each node in image data contains causal relationships with its neighboring nodes, neural networks based on Euclidean space cannot effectively learn the correlations between nodes. Therefore, Graph Convolutional Neural Networks (GCNs) are used to achieve intelligent diagnosis and status monitoring based on visual images.

[0047] Graph Neural Networks (GCNs) can deeply mine the positional relationships between image nodes, extract image data structural features and geometric relationships of key nodes, and define image features between given acquisition points. By extracting features from given image acquisition points and labeling the external features of the extracted basic units, the resulting image node feature data is used as input to the GCN model to obtain intelligent diagnostics of the device's status.

[0048] Furthermore, when feedback is received from the distributed multi-source contact diagnostic system and the mobile non-contact image diagnostic system, expert fuzzy fusion technology is used to provide task-related diagnostic method weights in real time, fuse contact and non-contact diagnostic results, obtain the operating status of mechanical equipment, classify and locate abnormal conditions, and realize the life prediction of mechanical equipment.

[0049] The beneficial effects of this invention are:

[0050] This invention proposes a technology based on distributed multi-source contact diagnostics and mobile non-contact image diagnostics. After initially optimizing the layout of the distributed multi-source sensing system and the mobile non-contact image system, the collected raw data is processed through a data fusion mechanism and an image analysis system to extract feature information. This extracted information is then input into DCADA and GNN network models to obtain intelligent monitoring and diagnosis of the mechanical equipment's status. This achieves a novel intelligent diagnostic method that combines signal and image information through distributed multi-source information fusion. By employing expert fuzzy fusion technology based on distributed multi-source contact diagnostics and mobile non-contact image diagnostics, comprehensive information monitoring of the mechanical equipment's operating status, classification and location of abnormal conditions, and prediction of the mechanical equipment's lifespan are achieved.

[0051] This invention presents a state-based intelligent monitoring and diagnostic method using contact signals and visual feedback. With its advantages of high diagnostic accuracy, comprehensive coverage, and low implementation difficulty, it can be widely used in production, daily life, and scientific research. Compared to diagnostic methods based solely on contact signal data or visual feedback, this integrated intelligent diagnostic method avoids the limitations and biases inherent in single-method approaches. The intelligent diagnostic method for the state of distributed multi-source sensing devices utilizes contact signal and visual feedback-based methods for mechanical equipment condition monitoring, intelligent diagnosis, and lifespan prediction, providing a new research approach and methodology for intelligent monitoring and diagnosis of mechanical equipment. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating a specific implementation of the present invention;

[0053] Figure 2 This is a diagram illustrating an embodiment of the multi-contact and non-contact sensor layout of the present invention;

[0054] Figure 3This is a schematic diagram of the contact sensor data fusion process of the present invention;

[0055] Figure 4 This is a schematic diagram of the deep convolutional adversarial domain adaptive neural network model of the present invention;

[0056] Figure 5 This is a diagram illustrating an embodiment of the multi-contact and non-contact sensor layout of the present invention. Specific implementation methods

[0057] To make the objectives and technical solutions of this invention clearer, the technical solutions and technical routes of this invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these examples are only used to illustrate the invention and are not intended to limit the scope of the invention. The described examples are some embodiments of this invention, but not all embodiments. All other embodiments obtained by those skilled in the art after reading this invention without creative effort are within the scope of protection of this invention.

[0058] This invention provides a method for intelligent diagnosis of device status using distributed multi-source sensing. Figure 1 A schematic flowchart of a distributed multi-source sensing device status intelligent diagnosis method provided as an example of the present invention is shown. The method includes:

[0059] like Figure 1 As shown, the distributed multi-source sensing intelligent equipment condition diagnosis method includes optimized layout of contact and non-contact sensors, contact signal acquisition system and non-contact image acquisition system, as well as DCADA and GCN neural network models. Through processes such as optimized arrangement of contact and non-contact sensors, acquisition of contact signal and image data, data preprocessing, intelligent diagnosis of network models, and expert fuzzy fusion of diagnostic information, the method achieves condition monitoring, intelligent diagnosis, and life prediction of mechanical equipment.

[0060] like Figure 1 and Figure 2 As shown, the distributed layout of contact sensors adopts the following steps: First, the kinematic relationship is analyzed based on the global coordinate system to establish a dynamic model of the mechanical equipment. Then, vibration, acceleration and strain sensors are arranged in the basic units such as the drive device, actuator, load-bearing component and transmission component of the mechanical equipment. In each electrically driven actuator, a signal detection channel is established by using voltage and current sensing modules.

[0061] For each basic unit, establish a rectangular coordinate system or cylindrical coordinate system that adapts to the shape and force characteristics of the unit. Construct an objective function for the distribution of sensors on the surface of the unit. Use the gradient descent method to solve for the optimal position and number of sensors distributed on the surface of the basic unit. For units that cannot be solved, arrange them in a centrally symmetrical manner with an odd number of sensors.

[0062] like Figure 1 and Figure 2 As shown, based on the dynamic model, kinematic relationships, and layout of contact sensors of the mechanical equipment, a non-contact image acquisition system is deployed using a tethered mobile platform. The tethered mobile platform includes, but is not limited to, a multi-rotor UAV, equipped with a binocular camera and a light source. Multiple image acquisition points are set up to assess the spatial layout of the mechanical equipment and the external representation of the motion of each basic unit, enabling real-time dynamic status monitoring and image capture. The optimal motion path is established using the A* algorithm.

[0063] like Figure 3 As shown, the contact signal analysis system uses raw data collected by contact sensors of mechanical equipment as input. It extracts signal features using Fast Fourier Transform (FFT) and Classical Mode Decomposition (MODE). Time-frequency domain features are extracted through signal processing techniques such as denoising, frequency reduction, decomposition, and compression. A multi-scale mechanism combines the extracted signal features and the original signal into three-channel data, achieving data fusion of distributed detection from multi-source contact sensors. FFT can effectively and quickly transform signals to obtain information about each frequency component in the frequency domain. By removing high-frequency signals, the filtered original signal is obtained. The specific process is as follows:

[0064] First, the expression for a discrete signal (Discrete Fourier Transform, DFT) of length N is given.

[0065]

[0066] To reduce computational complexity, a rotation factor is introduced.

[0067]

[0068] Decompose the signal into odd-even sequences.

[0069]

[0070] By utilizing periodicity and symmetry, signal sequences can be decomposed into odd and even sequences, thereby extracting the frequency characteristics of the signal.

[0071] Classical mode decomposition (CMD) is a time-frequency domain signal processing method that decomposes signals based on their inherent time-scale characteristics without requiring the setting of basis functions. It first decomposes the complex signal into a finite set of Intrinsic Mode Functions (IMFs). Each IMF component contains local feature information of the original signal at different time scales, and the signal's feature information is extracted through the IMFs. The classical mode decomposition process is as follows:

[0072] The input signal is decomposed into several IMFs and the residual r(t).

[0073]

[0074] By setting shutdown criteria, the IMF component c is obtained. i (t),c i (t) represents the component containing the maximum frequency of the input signal. The correlation coefficient p is calculated by substituting this component into the input signal. i When the correlation coefficient p i The closer it is to 1, the more the component c i (t) is closer to the original signal.

[0075]

[0076] Choose the correlation coefficient p i The highest IMF is used as the optimal mode function to preserve the time-frequency domain characteristics of the original signal.

[0077] The raw signals acquired by the contact sensor are processed through Fast Fourier Transform and Classical Mode Decomposition to form a three-channel signal data set, which is then combined with the original data. A length of N is extracted from each channel data set. 2 The continuous signal is converted into N×N two-dimensional image data, and grayscale processing is performed. Then, weight parameters a1, a2, and a3 are assigned to the three channels, where the relationship between a1, a2, and a3 is as follows:

[0078] a1 + a2 + a3 = 1, 0 ≤ a1, a2, a3 ≤ 1

[0079] Then, the grayscale images corresponding to the three channels are superimposed on the three-channel two-dimensional data images, and the contact-type multi-source sensing detection data fusion is achieved through a multi-scale weighting mechanism.

[0080] Environmental disturbances can easily distort data; fault data is significantly smaller than normal data in the collected data, causing data distribution shifts; and it is difficult to collect effective labeled data from sensors or to effectively label the data. Therefore, diagnostic results using deep learning and traditional signal processing methods are often unsatisfactory. To address these issues, this invention employs a deep convolutional adversarial domain adaptive neural network (DCADA) to achieve intelligent diagnosis of mechanical equipment under unsupervised, noisy, and data distribution shift conditions.

[0081] like Figure 4 As shown, the raw signal data collected by the contact sensor is preprocessed through data fusion, resulting in two domain sample data: source domain and target domain. The source domain sample data... Source domain tasks Includes labeled signal feature data, target domain x i ,xj vR d Target domain tasks These are signal feature data that do not contain labels, where n s ,n t Samples from the source and target domains, respectively, x i ,x j Representing the source domain sample data and the target domain sample data respectively, the feature spaces R of the two domains are... d They are the same.

[0082] The domain sample data obtained from the segmentation is used as input to the DCADA model, and its diagnostic model includes feature extraction, domain adaptation, and state recognition modules.

[0083] like Figure 4 The flowchart shown illustrates that the feature extraction module comprises five convolutional modules. Each convolutional module includes a convolutional layer, an activation function, a pooling layer, and finally a fully connected layer to extract features from the domain samples, converting the domain sample signals into learnable classification information. The loss function formula for its feature extraction module is as follows:

[0084]

[0085] The domain adaptation module helps the feature extraction module learn domain-invariant features by establishing a distribution difference metric and a multi-domain discriminator. Finally, a GRL gradient inversion layer backpropagates the loss function value to the feature extraction module. First, the features extracted by the feature extraction module are input into multiple domain discriminators, and the resulting probability values ​​are used as the L value of the domain discriminator. d The output is the loss value of the obtained domain discriminator as the marginal distribution difference of the domain. The inter-class distribution distance is calculated by using the maximum mean difference (MMD) and the resulting loss value is used as an adaptive factor to guide the domain adaptation of the marginal distribution difference.

[0086] To reduce the distributional differences among domain samples and extract domain-invariant features, multiple domain discriminators are used to classify source and target domain information. Each domain discriminator, acting as a standard binary classification network, includes global average pooling, convolutional blocks, and a softmax function module. The probability values ​​corresponding to the feature information are then used as the output of the domain discriminator, and its loss function is:

[0087]

[0088] Maximum Mean Difference (MMD) embeds high-level features from two domains into two points in the Regenerated Hilbert Space (RKHS) by introducing a mapping function. The distance between these two points represents the distributional distance between the features of the two domains. Matching the distributions of the two domains is achieved by minimizing this distance, where the distributional distance can be defined as...

[0089]

[0090] Where Φ is the mapping function of RKHS, X s With X t Let L represent samples from the source and target domains, respectively, and let L be the loss function of MMD. M :

[0091]

[0092] State recognition module: This module classifies features extracted adaptively from the domain to achieve anomaly classification and diagnosis. It includes convolutional blocks, fully connected layers, and a normalized softmax function. Its loss function is L... y for:

[0093]

[0094] The optimization objective for DCADA model parameter calculation is:

[0095]

[0096]

[0097] The iterative update formula for DCADA model parameters is:

[0098]

[0099]

[0100]

[0101] Non-contact image acquisition systems take non-contact acquired image data as input, extract image features of a given acquisition point through image processing techniques such as segmentation, denoising, enhancement, and grayscale, and label the external representations of the extracted basic units.

[0102] Due to the non-linear and irregular nature of image data, the nodes in an image vary in size, are unordered, and the relationships between adjacent nodes are difficult to determine. This makes it impossible to directly apply deep learning operations such as convolution to image data. Furthermore, since each node in an image data set contains causal relationships with its neighbors, neural networks based on Euclidean space cannot effectively learn these relationships. Therefore, graph neural networks are used to achieve state monitoring and intelligent diagnosis based on visual images.

[0103] Graph Neural Networks (GCNs) can deeply mine the positional relationships between image nodes, extract image data structural features and geometric relationships of key nodes, and define image features between given acquisition points. By extracting features from given image acquisition points and labeling the external features of the extracted basic units, the resulting image node feature data is used as input to the GCN model to obtain intelligent diagnostics of the device's status.

[0104] like Figure 1 As shown, based on the diagnostic results of the distributed multi-source contact diagnostic and mobile non-contact image diagnostic system, expert fuzzy fusion technology is used to obtain the operating status, classification and location of abnormal conditions of mechanical equipment, and to predict the life of mechanical equipment.

[0105] This invention proposes a novel intelligent diagnostic method based on distributed multi-source contact diagnostics and mobile non-contact image diagnostics. After initially optimizing the layout of the distributed multi-source sensing system and the mobile non-contact image system, the collected raw data is processed through a data fusion mechanism and an image analysis system to extract feature information. This extracted information is then input into DCADA and GCN network models to obtain intelligent diagnostics and condition monitoring of mechanical equipment. This achieves a novel intelligent diagnostic method that combines signal and image multi-source information fusion. By employing expert fuzzy fusion technology based on distributed multi-source contact diagnostics and mobile non-contact image diagnostics, the method enables the monitoring of the operating status of mechanical equipment, the classification and location of abnormal conditions, and the prediction of the equipment's lifespan.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications 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.

Claims

1. A method for intelligent diagnosis of device status using distributed multi-source sensing, characterized in that, Includes the following steps: Step 1: Based on the kinematic relationship, dynamic model and optimization objectives of the mechanical equipment, perform distributed optimization layout of multiple contact sensors to obtain distributed multi-source sensing feedback data; Step 2: To address the external representation of equipment motion, deploy a mobile dynamic image acquisition and analysis system to implement image-based non-contact equipment state extraction; Step 3: Using the distributed multi-source sensor feedback data from Step 1 as input, a data fusion mechanism is employed to obtain standard preprocessed feature data in the time-frequency domain; Step 4: Using the non-contact image data extracted in Step 2 as input, an image analysis system is used to extract key node features; Step 5: Using the feature data extracted in Step 3 as input to the network model, a deep convolutional adversarial domain adaptive neural network model is adopted to obtain the intelligent diagnostic results of the distributed multi-source contact equipment status. DCADA (Distributed Multi-Source Detection and Analysis) is used to obtain intelligent diagnostic results for the status of distributed multi-source contact devices. DCADA consists of feature extraction, domain adaptation, and status recognition modules. Based on distributed detection and fusion data from multi-source contact sensors, domain-invariant features and device status features are extracted to achieve intelligent diagnosis of device status. The loss function of the domain adaptation module of DCADA is: Among them, G f (x i G is the feature representation extracted by the feature extraction module. d (x) is the feature representation of the domain adaptive module, x i It is source domain sample data, y i It is source domain sample label data, d i Let W be the domain sample label of the i-th sample. c Here, n represents the weight coefficients of the discriminators, and n is the number of discriminators; the loss function of the state recognition module is: Among them, G y (x) is the feature representation of the state recognition module, n s The total number of samples in the source domain is given. The loss function values ​​of the domain adaptation module and the state recognition module are passed to the feature extraction module through the gradient reversal layer to obtain the iterative loss function of the feature extraction module. Where w is the marginal distribution weight of the maximum mean difference, θ f It is G f The weights are iteratively updated with parameters, where μ is the learning rate hyperparameter and L is the learning rate hyperparameter. M With L f These are the MMD marginal distribution loss function and the feature extraction module loss function, respectively. Step 6: Using the key node image features extracted in Step 4 as model input, a graph convolutional neural network model is adopted to obtain intelligent diagnosis of the mechanical equipment status. Step 7: Based on the intelligent diagnostic results of Step 5 and Step 6, expert fuzzy fusion technology is used to obtain the equipment's operating status, classification and location of abnormal situations, and to predict the equipment's lifespan.

2. The intelligent diagnostic method for device status using distributed multi-source sensing according to claim 1, characterized in that: The distributed layout of the contact sensors adopted is based on the kinematic relationship and dynamic model of the mechanical equipment. Vibration, acceleration and strain sensors are arranged on the drive device, actuator, load-bearing component and transmission component of the equipment. For each electric drive actuator, voltage and current sensing modules are used to establish an electrical signal detection channel.

3. The intelligent diagnostic method for device status using distributed multi-source sensing according to claim 2, characterized in that: A distributed multi-source sensor optimization layout is adopted. A rectangular coordinate system or cylindrical coordinate system adapted to the shape and force characteristics of each basic unit is established. An objective function for the distribution of sensors on the surface of the unit is constructed. The gradient descent method is used to solve for the optimal position and number of sensors distributed on the surface of the basic unit. For units that cannot be solved, a centrally symmetrical layout is adopted with an odd number of sensors.

4. The intelligent diagnostic method for device status using distributed multi-source sensing according to claim 1, characterized in that: A non-contact image acquisition system is deployed using a tethered mobile platform, which includes, but is not limited to, a multi-rotor UAV. Multiple image acquisition points are set up to consider the spatial layout of the equipment and the external representation of the motion of each basic unit, and the optimal motion path is established using the A* algorithm.

5. The intelligent diagnostic method for device status using distributed multi-source sensing according to claim 1, characterized in that: Contact-based multi-source sensor data fusion is adopted. For the raw data collected by the contact sensors, fast Fourier transform and classical mode decomposition are used to extract the signal features. The extracted signal features and the original signal are combined into three-channel data through a multi-scale fusion mechanism to realize the distributed detection data fusion of contact-based multi-source sensors.

6. The intelligent diagnostic method for device status using distributed multi-source sensing according to claim 1, characterized in that: Image analysis systems are used to obtain key node features. For non-contact image data, image features of a given acquisition point are extracted through image processing techniques such as segmentation, denoising, enhancement, and grayscale processing. The external representations of the extracted basic units are then labeled.

7. The intelligent diagnostic method for device status using distributed multi-source sensing according to claim 1, characterized in that: Intelligent diagnosis of equipment status is achieved by using key node image feature input. The GCN model is trained based on the multi-state of key nodes, and the newly acquired and processed node image data is input to obtain intelligent diagnosis of the mechanical device status.

8. The intelligent diagnostic method for device status using distributed multi-source sensing according to claim 1, characterized in that: By employing comprehensive intelligent diagnosis of mechanical equipment status, an expert fuzzy system is constructed to provide real-time weights of diagnostic methods relevant to the task. By fusing contact and non-contact diagnostic results, a fused equipment operating status, abnormal situation classification and location, and equipment life prediction are obtained.

Images