Multi-working-condition equipment health assessment method based on adaptive configuration convolution
By dynamically adjusting feature extraction through adaptive configuration convolutional networks, the problem of low feature extraction efficiency in traditional methods under multiple working conditions is solved, achieving efficient and accurate health assessment of mechanical equipment and adapting to feature capture under complex working conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING JIAOTONG UNIV
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional mechanical equipment health assessment methods perform poorly when dealing with the non-stationary, multi-scale characteristics of vibration signals under complex working conditions. They are unable to adaptively capture features at different scales, resulting in low feature extraction efficiency, increased computational overhead, and an inability to meet the adaptability and generalization requirements of actual industrial scenarios.
An adaptive configuration convolution method is adopted, which constructs an adaptive configuration convolution network through an adaptive configuration convolution module, a neural convergence module, and a neural divergence module. This dynamically adjusts the feature extraction process, reduces the number of computational parameters, adapts to vibration signal changes under various working conditions, and achieves efficient feature extraction.
It enables efficient and accurate health assessment of mechanical equipment under multiple operating conditions, reduces computational overhead, and improves the predictive and assessment performance of the model.
Smart Images
Figure CN122020120A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mechanical equipment health assessment, and specifically to a multi-condition equipment health assessment method based on adaptive configuration convolution. Background Technology
[0002] Mechanical equipment is a core component of modern industrial systems. From railways to aerospace, from heavy machinery to precision manufacturing, every critical link in modern industrial systems relies on the continuous and stable operation of mechanical equipment. The operating status and health level of mechanical equipment directly affect production safety, operational efficiency, and overall economic benefits. Therefore, conducting systematic health assessments of mechanical equipment is crucial. Based on scientific and accurate health assessments, maintenance personnel can rationally formulate maintenance plans and accurately determine maintenance timing, thereby avoiding unplanned downtime, reducing unnecessary maintenance expenses, and significantly improving the safety and stability of equipment operation.
[0003] Traditional health assessment methods largely rely on fixed-structure signal processing and feature extraction techniques, which perform poorly when dealing with the non-stationary and multi-scale characteristics of vibration signals under complex operating conditions, resulting in limited ability to identify early health assessments and progressive degradation processes. With the development of deep learning technology, especially the application of convolutional neural networks in time-series signal processing, new solutions have been provided for equipment health assessment. However, conventional convolutional neural network structures have limitations such as fixed receptive fields and rigid feature extraction methods, making it difficult to adapt to dynamic changes in signals and to adaptively capture features at different scales. This leads to low feature extraction efficiency, restricting its adaptability and generalization ability in practical industrial scenarios.
[0004] Improving model flexibility by giving convolution kernels the ability to dynamically adjust can effectively address the aforementioned problems. However, mechanical equipment often faces complex environments with varying operating conditions such as rotational speed and load during actual operation, resulting in strong non-stationarity and multi-scale characteristics in its vibration signals. This characteristic necessitates existing methods to learn a large number of parameters. Furthermore, because existing methods cannot achieve coordinated adaptive adjustment of the number and location of sampling points, they struggle to accurately capture key fault impact components while flexibly adapting to the significant scale differences in vibration characteristics under different operating conditions and degradation stages, based on the time-varying characteristics of the vibration signals. This structural limitation further exacerbates the computational burden, causing its cost to rise sharply with the increase in the number of sampling points.
[0005] Therefore, there is an urgent need for a mechanical equipment health assessment method that can dynamically optimize the feature extraction process under multiple working conditions, while taking into account both computational efficiency and model accuracy, so as to reduce the computational overhead caused by the increase in the number of sampling points and meet the dual requirements of efficiency and performance in practical engineering. Summary of the Invention
[0006] In view of this, the present invention provides a multi-condition equipment health assessment method based on adaptive configuration convolution to solve some of the technical problems in the background art.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] A multi-condition equipment health assessment method based on adaptive configuration convolution is designed, including the following steps: The vibration signals of mechanical equipment under various working conditions are acquired and normalized to obtain the original feature tensor. The original feature tensor is reduced in dimensionality using a standard convolution module to obtain a low-level data representation. An adaptive configuration convolutional network, constructed using an adaptive configuration convolutional module, a neural convergence module, and a neural divergence module, performs deep adaptive feature extraction on low-level data representations to obtain high-level data representations. The evaluation result output unit containing a fully connected layer is used to evaluate the high-level data representation and output the equipment health status evaluation result.
[0009] Furthermore, the adaptive configuration convolutional network performs deep adaptive feature extraction on the low-level data representation of the input, specifically expressed by the following formula:
[0010] This represents the high-level data representation of the output. AFB represents a low-level representation of the input data. ) represents an adaptive configuration convolutional module, D( ) represents the neural divergent module, C( ) represents the neural convergence module.
[0011] Furthermore, the adaptive configuration convolution module is represented by the following expression:
[0012] in To modify the activation function of the linear unit, For batch normalization layer, This represents the input features of the adaptive configuration convolutional module. This indicates that adaptive configuration convolution is applied to input features. The output features extracted from the above.
[0013] Furthermore, adaptive configuration convolution in input features The extracted output features are obtained through the following steps: Obtaining input features Left-side sampling features to the left of the sampling location And the right-side sampling feature to the right of the sampling position ; Calculate the sampling features on the left side respectively and right-side sampling features Weights:
[0014]
[0015] in, This indicates the interpolation weights of the left linear window. Indicates the interpolation weights of the right-hand linear window. Indicates the sampling location; This indicates the neighboring anchor point to the left of the sampling location. Indicates the neighboring anchor point to the right of the sampling location; Based on the sampling characteristics on the left right-side sampling features And the corresponding weighted linear window interpolation of the sampling position. :
[0016] Linear window interpolation based on sampling location This yields adaptive configuration convolutions in the input features. Output features extracted from above:
[0017]
[0018]
[0019] in, Y This represents the output features extracted by adaptive configuration convolution. This indicates that different convolution sizes are selected based on the features. This indicates element-wise multiplication. z This indicates the use of deformable convolutional tensor quantum networks to compute input features. The resulting deformable convolution tensor This indicates that the input features are calculated using a pivotal partial subnetwork. The obtained nucleus bias value, This represents the linear window interpolation at each sampling location. Indicates the kernel size as 3 Convolution operation, This indicates the Dropout layer. This represents the hyperbolic tangent function.
[0020] Furthermore, the sampling location is obtained in the following way: 1) Obtaining input features based on offset sensing base network offset sensing base P And based on the obtained offset sensing basis P The predicted dynamic sensing scale is obtained through a dynamic sensing scale network. l : Among them, the offset sensing base P Obtain it using the following formula:
[0021] in Indicates a modified linear unit. This represents a constant set to prevent division by zero. Describe the convolution operation. This represents the input features of the offset sensing base network. W P,1 and b P,1 These represent the weights and biases of the first convolutional layer in the offset-aware base network, respectively. W P,2 and b P,2 Let represent the weights and biases of the second convolutional layer in the offset-sensing base network, respectively. and These represent the offset and scaling parameters for the first batch of normalized layers, respectively. and These represent the offset and scaling parameters of the second batch of normalized layers, respectively. and These represent the mean and variance of the first normalized layer in a batch, respectively. and These represent the mean and variance of the second batch of normalized layers in a single batch, respectively. To modify the activation function of the linear unit;
[0022] in express function, W l,1 and b l,1 Let these represent the weight matrix and bias of the first convolutional layer of the variable receptive field network, respectively. W l,2 and b l,2 Let these represent the weight matrix and bias of the second convolutional layer of the variable receptive field network, respectively. and These represent the offset and scaling parameters for the first batch of normalized layers, respectively. and These represent the offset and scaling parameters of the second batch of normalized layers, respectively. and These represent the mean and variance of the first normalized layer in a batch, respectively. and These represent the mean and variance of the second batch of normalized layers in a single batch, respectively. 2) Generate the time-series control matrix U based on the number S of sampling points for the input features:
[0023] in S Indicates the number of sampling points. U The Middle i Column elements r i Defined as: ; 3) Construct the basic positional grid based on the outer product operation. :
[0024] Where T represents the number of sampling time steps. Indicates the outer product. Indicates the step size; 4) Based on dynamic sensing scale l The sampling position is obtained by using the time-controlled matrix U and the basic position grid.
[0025]
[0026] in This indicates element-wise multiplication. Represents dynamic sensing scale l The average value.
[0027] Furthermore, the number S of sampling points is obtained in the following way:
[0028] Where epoch represents the training epoch. This represents the number of sampling points calculated and saved when the training rounds reach 150. S , B Indicates the batch sample size. T Indicates the number of time steps. express Round down; Defined as:
[0029] in This represents Iverson brackets, indicating that if the condition inside the brackets is satisfied... y If the number is odd, the value is 1; otherwise, the value is 0. Represents the maximum value based on a given time scale. Determined time-series adjustment factor.
[0030] Furthermore, the left neighbor anchor point of each sampling location and right neighbor anchor Obtain them through the following steps:
[0031]
[0032] in Indicates the left neighbor anchor point. Indicates the right neighbor anchor point. express function, Indicates the sampling position Round down. T Indicates the number of time steps.
[0033] Furthermore, the neural convergence module specifically includes the following expression:
[0034] in Indicates the kernel size as k Convolution operation, This represents the input features of the neural convergence module. Indicates the training round.
[0035] Furthermore, the neural divergence module specifically includes the following expression:
[0036] in, and This represents two input features of the neural divergence module. Indicates the kernel size as k The transpose of the matrix, Indicates the training round.
[0037] Furthermore, in the step of evaluating the high-level data representation using an evaluation result output unit containing a fully connected layer and outputting the equipment health status evaluation result, the evaluation result output unit outputs the equipment health status evaluation result, which is expressed by the following expression:
[0038] in This represents a high-level data representation of the output of an adaptive configuration convolutional network. Indicates a fully connected layer. For flattening operation, This indicates global average pooling.
[0039] As can be seen from the above technical solution, this invention discloses a multi-condition equipment health assessment method based on adaptive configuration convolution, which has the following beneficial effects compared with the prior art: The method of this invention constructs an adaptive configuration convolution, which can dynamically determine the number of features to be extracted based on multi-condition data, thereby adjusting the spatial distribution of the extracted features and the sampling window, reducing the number of computational parameters, and has good prediction and evaluation performance. Attached Figure Description
[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0041] Figure 1 A flowchart of the multi-condition equipment health assessment method based on adaptive configuration convolution provided by the present invention; Figure 2 This is a framework diagram of adaptive configuration convolution provided by the present invention; Figure 3 The flowchart provided by this invention is for constructing a model based on adaptive configuration convolution. Figure 4 This is a performance comparison between the method of the present invention and existing methods. Detailed Implementation
[0042] The technical solutions of the embodiments 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, and 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.
[0043] This invention discloses a multi-condition equipment health assessment method using adaptive configuration convolution, comprising the following steps: The vibration signals of mechanical equipment under various working conditions are acquired and normalized to obtain the original feature tensor. The original feature tensor is reduced in dimensionality using a standard convolution module to obtain a low-level data representation. An adaptive configuration convolutional network, constructed using an adaptive configuration convolutional module, a neural convergence module, and a neural divergence module, performs deep adaptive feature extraction on low-level data representations to obtain high-level data representations. The evaluation result output unit containing a fully connected layer is used to evaluate the high-level data representation and output the equipment health status evaluation result.
[0044] In this invention, a dataset is first obtained from real-time vibration signals during the operation of mechanical equipment. Any dataset can be represented as... ,in Indicates the first i A sample of real-time vibration signal data. This represents the corresponding label. The dataset is normalized to obtain the data. ,in Indicates the first training samples, This indicates the corresponding tag, where Where 0 represents a complete failure state and 1 represents a complete health state; the normalized expression is:
[0045] in, express The mean, N Indicates the number of samples.
[0046] Then, the normalized original feature tensor is input into the health assessment model for assessment and prediction. To facilitate understanding of the overall process of the health assessment model, it is described in three parts.
[0047] The first part consists only of standard convolutions and is used to preprocess the input data. By reducing the dimensionality of the data, it transforms the raw sensor data into a low-level data representation. The first part consists of the following layers:
[0048] in This represents the low-level data representation of the first part of the output. Indicates the number of output channels. Indicates the kernel size as 3 The convolution operation.
[0049] The second part consists of an adaptive configuration convolutional module, a neural convergence module, and a neural divergence module stacked together. This increases the depth of the network model, captures short-term abnormal features and long-term trend evolution of mechanical equipment degradation under various operating conditions, and achieves more accurate health assessment. The second part comprises the following layers:
[0050] in This represents the high-level data representation of the second part of the output, AFB ( ) represents an adaptive configuration convolutional module, D( ) represents the neural divergent module, C( ) represents the neural convergence module.
[0051] The third part consists of unfolded layers and linear layers, used to receive high-level representations extracted by the preceding network. The health status of the mechanical equipment is assessed. Part Three consists of the following levels:
[0052] in Indicates a fully connected layer. For flattening operation, This indicates global average pooling.
[0053] The constructed adaptive configuration convolutional network is trained iteratively until the iteration task ends, and the health status of the mechanical equipment is output.
[0054] The core of this invention lies in the construction of an adaptive configuration convolution module. The principles of this invention will be described in detail below through the specific structure and data processing flow of the adaptive configuration convolution.
[0055] The adaptive configuration convolution module disclosed in this invention has the following characteristics: The offset sensing basis of the input feature tensor is calculated by an offset sensing basis network, and the dynamic sensing scale is predicted based on the offset sensing basis by a dynamic sensing scale network. The number of sampling points is dynamically adjusted based on the number of training cycles and a given time scale to achieve adaptive optimization of the sampling strategy during training. A temporal control matrix and a basic position grid are generated based on the number of sampling points, and the sampling positions of the final adaptive configuration convolution are calculated in combination with the dynamic sensing scale. By calculating the neighborhood anchor points of the sampling location and applying boundary constraints, a sampling window for adaptive configuration convolution is constructed. Calculate linear window interpolation, select convolution kernels of corresponding sizes from multiple convolution kernel size libraries for convolution operation, and perform affine transformation through pivot kernel bias and deformable convolution tensor to obtain adaptive configuration convolution output features.
[0056] The adaptive configuration convolution with the above characteristics is incorporated into the health assessment model, and iterative training is performed until the end of the iteration task, outputting the equipment health status.
[0057] The following section provides a further explanation of the top-level adaptive configuration convolutional module in the aforementioned health assessment model. The specific process is as follows: Figures 1-2 As shown.
[0058] First, the offset sensing base is calculated using the offset sensing base network. P Input features are amplified to enhance gradients. Offset-sensor base. P The calculation method is as follows:
[0059] in Indicates a modified linear unit. Indicates the negative slope parameter. This represents a very small constant set to prevent division by zero. Describe the convolution operation. x Indicates input features, W P,1 and b P,1 These represent the weights and biases of the first convolutional layer in the offset-aware base network, respectively. W P,2 and b P,2 Let represent the weights and biases of the second convolutional layer in the offset-sensing base network, respectively. and These represent the offset and scaling parameters for the first batch of normalized layers, respectively. and These represent the offset and scaling parameters of the second batch of normalized layers, respectively. and These represent the mean and variance of the first normalized layer in a batch, respectively. and These represent the mean and variance of the second batch of normalized layers in a single batch, respectively. To modify the activation function of the linear unit.
[0060] Subsequently, through a dynamic sensing scale network, based on the offset sensing base... P Predict the dynamic sensing scale at each time point l Dynamic sensing scale l The calculation method is as follows:
[0061] in express function, W l,1 and b l,1 Let these represent the weight matrix and bias of the first convolutional layer of the variable receptive field network, respectively. W l,2and b l,2 Let these represent the weight matrix and bias of the second convolutional layer of the variable receptive field network, respectively. and These represent the offset and scaling parameters for the first batch of normalized layers, respectively. and These represent the offset and scaling parameters of the second batch of normalized layers, respectively. and These represent the mean and variance of the first normalized layer in a batch, respectively. and These represent the mean and variance of the second batch of normalized layers in a single batch, respectively.
[0062] The process of dynamically adjusting the number of sampling points based on the number of training epochs and a given time scale specifically includes: S31. When the given time scale range is Time series adjustment factor Set to 1; otherwise, it will be based on the maximum value of the given time scale. Determine the time series adjustment factor The specific method is as follows:
[0063] in t Indicates a given time scale. otherwise Representing an interval Other than these.
[0064] S32. Dynamically adjust the number of sampling points based on the number of training cycles. When the number of training cycles is less than 150, use a timing adjustment factor. and dynamic sensing scale l Determine the number of sampling points S When the number of training rounds exceeds 150, the number of sampling points... S Using the number of sampling points from round 150 The specific method is as follows:
[0065] Where epoch represents the training epoch. This represents the number of sampling points calculated and saved when the training rounds reach 150. S , B Indicates the batch sample size. T Indicates the number of time steps. express Round down; Defined as:
[0066] in This represents Iverson brackets, indicating that if the condition inside the brackets is satisfied... x If the number is odd, the value is 1; otherwise, the value is 0.
[0067] Next, a temporal adjustment matrix and a basic position grid are generated based on the number of sampling points. Combined with the dynamic sensing scale, the sampling positions for the final adaptive configuration convolution can be calculated. The specific process is as follows: S41. Number of sampling points used S Generate timing regulation matrix U Timing control matrix U Defined as:
[0068] in S Indicates the number of sampling points. U The Middle i Column elements r i Defined as:
[0069] S42. Constructing a basic location grid using outer product operations. The specific method is as follows:
[0070] in Indicates the outer product. Indicates the step size.
[0071] S43. Generate a basic position grid based on the number of sampling points, and combine the basic position grid and the dynamic perception scale to generate the final sampling position of the adaptive configuration convolution. The specific method is as follows:
[0072] in This indicates element-wise multiplication. Represents dynamic sensing scale l The average value.
[0073] After obtaining the sampling positions, a sampling window for adaptive configuration convolution is constructed by calculating the neighborhood anchor points of the sampling positions. Neighborhood anchor points and The calculation formulas are as follows:
[0074]
[0075] in Indicates the left neighbor anchor point. Indicates the right neighbor anchor point. express function, Indicates to Round down. T Indicates the number of time steps.
[0076] After obtaining the neighborhood anchor points of the sampling window, linear window interpolation can be calculated, and affine transformation can be performed using the pivot kernel bias and deformable convolution tensor to obtain the adaptive configuration convolution output features. The specific process is as follows: S61. Calculate the linear window interpolation weights for each sampling position to obtain the left and right linear window interpolation weights. and The calculation methods are as follows:
[0077]
[0078] in This indicates the interpolation weights of the left linear window. This represents the interpolation weight of the right-hand linear window.
[0079] S62. Extract sampling features of the left and right sides from the input features. and The linear window interpolation is calculated using linear window interpolation weights, as follows:
[0080] in Indicates the sampling features on the right. This indicates the sampling features on the left.
[0081] S63. Select convolution kernels of corresponding sizes from multiple convolution kernel size libraries to perform convolution operations on linear window interpolation, and use two sub-networks to calculate the pivot kernel bias. and deformable convolution tensor z Perform an affine transformation on the output of the convolution operation to generate the output features of the adaptive configuration convolution. Y The specific method is as follows:
[0082] in This indicates that different convolution sizes are selected based on the features. Y This represents the output features of adaptive configuration convolution. This indicates element-wise multiplication.
[0083] The above process illustrates the data processing method of the top-level adaptive configuration convolution in the health assessment model. The data processing of other layers of adaptive configuration convolution modules in the health assessment model is similar to the above process, the difference being the input features. The overall flow of the health assessment model is as follows:
[0084] The specific health assessment process is as follows: An adaptive configuration convolutional module is embedded in the model to dynamically adjust the number, location, and size of samples to adapt to complex time-varying patterns under multiple operating conditions, enabling health assessment of mechanical equipment under various operating conditions. This module consists of the following layers:
[0085] in To modify the activation function of the linear unit, For batch normalization layer, This represents the input features of the adaptive configuration convolutional module. This indicates that adaptive configuration convolution is applied to input features. The output features extracted from the above.
[0086] Constructing a neural convergence module , The neural convergence module represents the input features, and its structure changes with each training round. Specifically, the neural convergence module consists of the following layers:
[0087] in Indicates the kernel size as k × k The convolution operation.
[0088] Constructing a neural divergence module , and The neural divergent module represents two input features, and its structure changes with each training round. Specifically, the neural divergent module consists of the following layers:
[0089] like Figure 3 As shown, an adaptive configuration convolutional network is constructed using an adaptive configuration convolutional module, a neural convergence module, and a neural divergence module to perform health assessment on mechanical equipment. Here, the output function of the adaptive configuration convolutional network is defined as follows: Given data points The adaptive configuration convolution output function can be expressed as: ,in WIndicates the width of the feature. Indicates the number of input channels. This represents the parameters of an adaptive configuration convolutional network.
[0090] This experiment selected rolling bearings as the research object and designed and carried out a health status assessment experiment for mechanical equipment under multiple working conditions. The bearing model used in the experiment was LDK UER204, and its specific parameters are shown in Table 1. The experimental platform consisted of an AC motor, a rotating shaft, a motor speed controller, an acceleration sensor, a hydraulic loading system, and test bearings, etc., used to collect bearing vibration data.
[0091] Table 1. Bearing-related parameters
[0092] The platform's adjustable operating parameters mainly include radial force and rotational speed, as shown in Table 2. The radial force is applied to the bearing housing of the test bearing via a hydraulic loading system, while the rotational speed is set and adjusted by the AC motor's speed controller. Three different operating conditions were set up for the experiment, each containing five bearing samples. Vibration signals were collected from each bearing in both horizontal and vertical directions. During data acquisition, the sampling frequency was set to 25.6 kHz, with one sample taken per minute and each sample lasting 1.28 seconds.
[0093] Table 2. Different Operating Conditions in the Design
[0094] The experiment uses the method proposed in this invention to perform health assessment on rolling bearings in mechanical equipment under multiple working conditions, and compares it with three commonly used methods: Content-Adaptive Non-Local Convolution (CANConv), Pyramidal Convolution (PyConv), and Omni-Dimensional Dynamic Convolution (ODConv). To facilitate comparative experiments, this section uses Cumulative Relative Accuracy (CRA) and Root Mean Square Error (RMSE) to quantitatively evaluate performance, and trains all three methods using the same hyperparameters and network structure. The relevant training parameters are shown in Table 3. Experimental results are summarized in Table 4 and... Figure 4 .
[0095] Table 3. Summary of Training-Related Parameters
[0096] Table 4. Summary of Comparison Results with Existing Methods
[0097] The experimental results show that the method of this invention exhibits superior prediction performance compared to the other three methods.
[0098] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0099] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A multi-condition equipment health assessment method based on adaptive configuration convolution, characterized in that, Includes the following steps: The vibration signals of mechanical equipment under various working conditions are acquired and normalized to obtain the original feature tensor. The original feature tensor is reduced in dimensionality using a standard convolution module to obtain a low-level data representation. An adaptive configuration convolutional network, constructed using an adaptive configuration convolutional module, a neural convergence module, and a neural divergence module, performs deep adaptive feature extraction on low-level data representations to obtain high-level data representations. The evaluation result output unit containing a fully connected layer is used to evaluate the high-level data representation and output the equipment health status evaluation result.
2. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 1, characterized in that, Adaptive configuration convolutional networks perform deep adaptive feature extraction on low-level data representations of the input, specifically expressed by the following formula: This represents the high-level data representation of the output. AFB represents a low-level representation of the input data. ) represents an adaptive configuration convolutional module, D( ) represents the neural divergent module, C( ) represents the neural convergence module.
3. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 2, characterized in that, The adaptive configuration convolution module is represented by the following expression: in To modify the activation function of the linear unit, For batch normalization layer, This represents the input features of the adaptive configuration convolutional module. This indicates that adaptive configuration convolution is applied to input features. The output features extracted from the above.
4. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 3, characterized in that, Adaptive configuration convolution in input features The extracted output features are obtained through the following steps: Obtaining input features Left-side sampling features to the left of the sampling location And the right-side sampling feature to the right of the sampling position ; Calculate the sampling features on the left side respectively and right-side sampling features Weights: in, This indicates the interpolation weights of the left linear window. Indicates the interpolation weights of the right-hand linear window. Indicates the sampling location; This indicates the neighboring anchor point to the left of the sampling location. This indicates the neighboring anchor point to the right of the sampling location; Based on the sampling characteristics on the left Right side sampling features And the corresponding weighted linear window interpolation at the sampling location. : Linear window interpolation based on sampling location This yields adaptive configuration convolutions in the input features. Output features extracted from above: in, Y This represents the output features extracted by adaptive configuration convolution. This indicates that different convolution sizes are selected based on the features. This indicates element-wise multiplication. z This indicates the use of deformable convolutional tensor quantum networks to compute input features. The resulting deformable convolution tensor This indicates that the input features are calculated using a pivotal partial subnetwork. The obtained nucleus bias value, This represents the linear window interpolation at each sampling location. Indicates the kernel size as 3 Convolution operation, This indicates the Dropout layer. This represents the hyperbolic tangent function.
5. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 4, characterized in that, The sampling location is obtained in the following way: 1) Obtaining input features based on offset sensing base network offset sensing base P And based on the obtained offset sensing basis P The predicted dynamic sensing scale is obtained through a dynamic sensing scale network. l : Among them, the offset sensing base P Obtain it using the following formula: in Indicates a corrected linear unit. This represents a constant set to prevent division by zero. Describe the convolution operation. This represents the input features of the offset sensing base network. W P,1 and b P,1 These represent the weights and biases of the first convolutional layer in the offset-aware base network, respectively. W P,2 and b P,2 Let represent the weights and biases of the second convolutional layer in the offset-sensing base network, respectively. and These represent the offset and scaling parameters for the first batch of normalized layers, respectively. and These represent the offset and scaling parameters of the second batch of normalized layers, respectively. and These represent the mean and variance of the first normalized layer in a batch, respectively. and These represent the mean and variance of the second batch of normalized layers in a single batch, respectively. To modify the activation function of the linear unit; in express function, W l,1 and b l,1 Let these represent the weight matrix and bias of the first convolutional layer of the variable receptive field network, respectively. W l,2 and b l,2 Let these represent the weight matrix and bias of the second convolutional layer of the variable receptive field network, respectively. and These represent the offset and scaling parameters for the first batch of normalized layers, respectively. and These represent the offset and scaling parameters of the second batch of normalized layers, respectively. and These represent the mean and variance of the first normalized layer in a batch, respectively. and These represent the mean and variance of the second batch of normalized layers in a single batch, respectively. 2) Generate the time-series control matrix U based on the number S of sampling points for the input features: in S Indicates the number of sampling points. U The Middle i Column elements r i Defined as: ； 3) Construct the basic positional grid based on the outer product operation. : Where T represents the number of sampling time steps. Indicates the outer product. Indicates the step size; 4) Based on dynamic sensing scale l The sampling position is obtained by using the time-controlled matrix U and the basic position grid. in This indicates element-wise multiplication. Represents dynamic sensing scale l The average value.
6. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 5, characterized in that, The number of sampling points S is obtained in the following way: Where epoch represents the training epoch. This represents the number of sampling points calculated and saved when the training rounds reach 150. S , B Indicates the batch sample size. T Indicates the number of time steps. express Round down; Defined as: in This represents Iverson brackets, indicating that if the condition inside the brackets is satisfied... y If the number is odd, the value is 1; otherwise, the value is 0. Represents the maximum value based on a given time scale. Determined time-series adjustment factor.
7. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 4, characterized in that, Left neighbor anchor point of each sampling location and right neighbor anchor Obtain them through the following steps: in Indicates the left neighbor anchor point. Indicates the right neighbor anchor point. express function, Indicates the sampling position Round down. T Indicates the number of time steps.
8. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 2, characterized in that, The neural convergence module specifically includes the following expressions: in Indicates the kernel size as k Convolution operation, This represents the input features of the neural convergence module. Indicates the training round.
9. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 2, characterized in that, The neural divergence module specifically includes the following expressions: in, and This represents two input features of the neural divergent module. Indicates the kernel size as k The transpose of the matrix, Indicates the training round.
10. In the multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 1, in the step of evaluating high-level data representation using an evaluation result output unit containing a fully connected layer and outputting equipment health status assessment results, the evaluation result output unit outputs the equipment health status assessment results, which are expressed by the following expression: in This represents a high-level data representation of the output of an adaptive configuration convolutional network. Indicates a fully connected layer. For flattening operation, This indicates global average pooling.