Special vehicle key component health assessment method based on orthogonal characteristic dynamic coupling

Abstract

The invention discloses a special vehicle key component health assessment method based on orthogonal feature dynamic coupling, which comprises the following steps: respectively carrying out feature extraction on a horizontal vibration signal and a vertical vibration signal in an initial input tensor based on a homogeneous convolution kernel to obtain a corresponding horizontal unidirectional feature and a corresponding vertical unidirectional feature; performing feature extraction on the initial input tensor based on a heterogeneous convolution kernel to obtain orthogonal direction features; external knowledge is introduced, and coupling processing is carried out on the features; according to the final coupling feature, combining a health degree label to obtain a final output health degree; and finally, obtaining an optimal health assessment model through iterative optimization training according to the loss function. And based on the optimal health assessment model, health assessment of the key component of the target special vehicle is realized. According to the method, the dynamic correlation and the coupling relationship between the signal features in different directions can be mined, the accuracy of health assessment is improved, and the generalization of the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of health assessment of key components of special vehicles, and more specifically to a health assessment method of key components of special vehicles based on orthogonal feature dynamic coupling. Background Art

[0002] Components such as bearings and gears are widely used in key systems such as special vehicle power energy and electromechanical transmission. Their health status is directly related to the vehicle's operating reliability and service life. By accurately evaluating the health status of key components of special vehicles, potential faults can be discovered in advance and precise situation-based maintenance strategies can be formulated to improve the availability and safety of the vehicle and ensure the smooth completion of the mission. In recent years, thanks to the rapid development of sensor technology and the improvement of the computing power of on-board equipment, health assessment methods for key components of special vehicles based on artificial intelligence have gradually been applied. These methods usually use vibration signals as model inputs, and then use convolutional neural networks, deep belief networks, recurrent neural networks, etc. to automatically learn signal features and conduct health assessments, which has become a strong support for ensuring the safe and reliable operation of special vehicles.

[0003] Due to the complex operating environment and time-varying service conditions of special vehicles, in order to obtain more comprehensive health status information of key components of special vehicles, vibration signals in two orthogonal directions, vertical and horizontal, are usually collected. These two orthogonal vibration signals can not only reflect the changes in vibration amplitude in different directions, but also characterize the phase change relationship between different directions, thus more fully describing the service performance degradation behavior of key components of special vehicles.

[0004] However, although the existing AI-based health assessment method for key components of special vehicles uses two orthogonal vibration signals as input, it only focuses on the extraction of single-direction signal features during feature learning, ignoring the dynamic correlation and coupling relationship between signal features in different directions, resulting in incomplete extracted health status information, which in turn affects the accuracy of health assessment of key components of special vehicles. In addition, the existing AI-based health assessment method for key components of special vehicles usually relies on a pre-defined model architecture for signal feature extraction and learning, and this fixed model architecture is difficult to dynamically adapt to complex operating conditions and a wide variety of failure modes, resulting in poor generalization performance of existing health assessment methods in actual engineering applications, making them difficult to promote and use on a large scale.

[0005] Therefore, how to explore the dynamic correlation and coupling relationship between signal features in different directions, improve the accuracy of health assessment, and how to improve the generalization performance of the model in the face of complex working conditions and diverse failure modes are issues that technical personnel in this field urgently need to solve. Summary of the invention

[0006] In view of the above problems, the present invention provides a health assessment method for key components of special vehicles based on orthogonal feature dynamic coupling to at least solve some of the technical problems mentioned in the above background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] The present invention provides a health assessment method for key components of special vehicles based on orthogonal feature dynamic coupling, comprising the following steps:

[0009] Acquire a vibration signal data set of the vehicle-mounted equipment under different working conditions; the vibration signal data set includes two orthogonal direction vibration signals of multiple samples and corresponding health labels;

[0010] Batch loading the vibration signal data set, and converting the samples therein into a tensor form after preprocessing to obtain an initial input tensor; which is used as the input of the initial health assessment model;

[0011] Based on the homogeneous convolution kernel, the horizontal vibration signal and the vertical vibration signal in the initial input tensor are respectively extracted to obtain the corresponding horizontal single-direction feature and vertical single-direction feature;

[0012] Performing feature extraction on the initial input tensor based on a heterogeneous convolution kernel to obtain orthogonal direction features;

[0013] By introducing external knowledge, coupling processing is performed on the horizontal unidirectional feature, the vertical unidirectional feature and the orthogonal directional feature to obtain a final coupling feature;

[0014] According to the final coupling feature, combined with the health label, a final output health is obtained;

[0015] Calculate a loss function according to the final output health and the health label;

[0016] According to the loss function, the initial health assessment model is iteratively optimized and trained to obtain an optimal health assessment model;

[0017] Based on the optimal health assessment model, health assessment of key components of target special vehicles is achieved.

[0018] Furthermore, the vibration signal data set is batch loaded, and two orthogonal vibration signals are preprocessed to obtain an initial input tensor; specifically, the method includes:

[0019] The vibration signal dataset is represented as ;in, Indicates horizontal vibration signal; Indicates vertical vibration signal; represents the health label corresponding to the i-th sample; N represents the total number of samples; the vibration signals in two orthogonal directions of the i-th sample are expressed as ;

[0020] The vibration signal dataset Divide into multiple batches for loading, and the vibration signal data of each batch is represented as , where b means there are b samples in each batch;

[0021] Each sample in each batch is normalized and converted into a tensor form to obtain the initial input tensor; the initial input tensor is expressed as .

[0022] Furthermore, based on the homogeneous convolution kernel, feature extraction is performed on the horizontal vibration signal in the initial input tensor to obtain the corresponding horizontal unidirectional feature; specifically, the process includes:

[0023] (1) The horizontal vibration signal is processed through the first-layer homogeneous convolution kernel and the corresponding weights. Convolution, normalization and activation operations are performed to obtain the horizontal first layer single-directional features, which can be expressed as:

[0024]

[0025] in, Represents the single-directional features of the first horizontal layer; represents the activation function; represents the normalization operation; Conv(·) represents the convolution process; Represents the first-layer homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0026] (2) Through the second-layer homogeneous convolution kernel and corresponding weights, the single-directional features of the first-layer After convolution, normalization and activation operations, residual connection is performed to obtain the horizontal second layer single-direction features, which can be expressed as:

[0027]

[0028] in, Represents the single-directional features of the horizontal second layer; Represents the second-layer homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0029] (3) Through the third-layer homogeneous convolution kernel and corresponding weights, the single-directional features of the second-layer After convolution, normalization and activation operations, residual connection is performed to obtain the horizontal third layer unidirectional features as the final horizontal unidirectional features; it is expressed as:

[0030]

[0031] Represents horizontal unidirectional features; Represents the third-layer homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term.

[0032] Furthermore, based on the homogeneous convolution kernel, feature extraction is performed on the vertical vibration signal in the initial input tensor to obtain the corresponding vertical unidirectional feature; specifically, the process includes:

[0033] (1) The vertical vibration signal is processed through the vertical first layer homogeneous convolution kernel and corresponding weights Convolution, normalization and activation operations are performed to obtain the vertical first-layer single-direction features, which can be expressed as:

[0034]

[0035] in, Indicates the vertical first layer unidirectional characteristics; represents the activation function; represents the normalization operation; Conv(·) represents the convolution process; Represents the first vertical homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0036] (2) Through the vertical second layer homogeneous convolution kernel and corresponding weights, the single-directional features of the vertical first layer After convolution, normalization and activation function, residual connection is performed to obtain the vertical second layer single-direction features, which can be expressed as:

[0037]

[0038] in, Indicates the vertical second layer unidirectional characteristics; Represents the second vertical layer of homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0039] (3) Through the vertical third layer homogeneous convolution kernel and corresponding weights, the vertical second layer single-directional features After convolution, normalization and activation operations, residual connection is performed to obtain the vertical third layer unidirectional features as the final vertical unidirectional features; it is expressed as:

[0040]

[0041] in, Indicates vertical unidirectional features; Represents the third vertical layer of homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term.

[0042] Furthermore, the feature extraction of the initial input tensor based on the heterogeneous convolution kernel to obtain orthogonal direction features specifically includes:

[0043] (1) Through the first layer of heterogeneous convolution kernels and corresponding weights, the initial input tensor is convolved, normalized and activated to obtain the first layer output features; expressed as:

[0044]

[0045] in, Represents the first layer output features; represents the activation function; represents the normalization operation; Conv(·) represents the convolution process; Represents the first layer of heterogeneous convolution kernel; express The corresponding weights; express The corresponding bias term; represents the initial input tensor;

[0046] (2) Through the second layer of heterogeneous convolution kernels and corresponding weights, the first layer output features After convolution, normalization and activation operations, residual connection is performed to obtain the second layer output features; it is expressed as:

[0047]

[0048] in, Represents the second layer output features; Represents the second layer of heterogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0049] (3) Through the third layer of heterogeneous convolution kernels and corresponding weights, the second layer output features After convolution, normalization and activation operations, residual connection is performed to obtain the output features of the third layer, which are used as the final orthogonal direction features; it is expressed as:

[0050]

[0051] in, Represents orthogonal directional features; Represents the third layer of heterogeneous convolution kernel; express The corresponding weights; express The corresponding bias term.

[0052] Furthermore, the horizontal unidirectional feature, the vertical unidirectional feature and the orthogonal directional feature are coupled by introducing external knowledge to obtain the final coupled feature; specifically including:

[0053] (1) The horizontal single-direction feature, the vertical single-direction feature, and the orthogonal direction feature are spliced ​​to obtain a multi-scale splicing feature; expressed as:

[0054]

[0055] Among them, F fusion input Represents multi-scale splicing features; Concat(·) represents the splicing operation; Represents horizontal unidirectional features; Indicates vertical unidirectional features; Represents orthogonal directional features;

[0056] (2) Encoding the external knowledge to obtain external knowledge encoding features; the external knowledge includes historical health data, equipment characteristics, and fault thresholds of the vehicle-mounted equipment; the external knowledge encoding features are expressed as:

[0057]

[0058] K external represents the external knowledge encoding feature; Encode(·) represents the encoding operation; H represents the historical health data; P represents the equipment characteristics; T represents the fault threshold; represents the external weight matrix;

[0059] (3) Perform nonlinear mapping on the external knowledge encoding features to obtain high-dimensional features; expressed as:

[0060]

[0061] K transformer Represents high-dimensional features; represents the activation function; Represents the weight of nonlinear mapping; Represents the bias of the nonlinear mapping;

[0062] (4) Multi-scale splicing feature F fusion input and high-dimensional features K transformer Perform fusion processing to obtain the fusion feature F enhanced ; expressed as:

[0063]

[0064] in, and Both represent weights adaptively adjusted by feature importance, specifically:

[0065]

[0066] Among them, Attention(·) represents the self-attention operation;

[0067] (5) Fusion feature F enhanced Perform global pooling to obtain global features; expressed as:

[0068]

[0069] Among them, F global Represents global features; GlobalAvgPool(·) represents the pooling layer operation;

[0070] (6) Obtain global features F through the fully connected layer global The feature weights are normalized using the Softmax function to obtain a weight vector containing all global features; it is expressed as:

[0071]

[0072] Where w represents the weight vector containing all features; Linear(·) represents the fully connected layer operation;

[0073] (7) Based on the weight vector w, the multi-scale splicing feature F fusion input Perform coupling processing to obtain the initial coupling characteristics; where the weight vector w includes the horizontal unidirectional characteristics Weight , vertical single direction feature The weight w2 and orthogonal direction features Weight ,and ;

[0074] The initial coupling characteristic F combined It is expressed as:

[0075]

[0076] (8) Multi-scale splicing feature F fusion input and the initial coupling feature F combined Perform residual connection to obtain residual features; expressed as:

[0077]

[0078] Among them, F final represents the residual feature;

[0079] (9) The residual features are transformed nonlinearly through the activation function to obtain the final coupling features; expressed as:

[0080]

[0081] Among them, F final output represents the final coupled feature.

[0082] Furthermore, the final output health is obtained according to the final coupling feature and in combination with the health label; specifically including:

[0083] (1) Based on the final coupling feature, the unidirectional feature is decoded to obtain the unidirectional decoding feature; it is expressed as:

[0084]

[0085] Among them, R homo Represents unidirectional decoding features; represents the activation function; W homo represents the unidirectional decoding weight; b homo Indicates unidirectional decoding bias; F final output represents the final coupling characteristics;

[0086] (2) Based on the final coupling feature, the orthogonal direction feature is decoded to obtain the orthogonal direction decoding feature; which is expressed as:

[0087]

[0088] Among them, R hetero represents the orthogonal direction decoding feature; W hetero represents the decoding weight in the orthogonal direction; b hetero Indicates the decoding bias in the orthogonal direction;

[0089] (3) Decoding feature R in one directionhomo And the orthogonal direction decoding feature R hetero Perform dynamic coupling to obtain the initial output health R output ; expressed as:

[0090]

[0091] in, Indicates dynamic adjustment of weights, specifically:

[0092]

[0093] Where Var(·) represents variance;

[0094] (4) Mapping the initial output health into a label distribution of a time series to obtain a health value distribution; expressed as:

[0095]

[0096] in, represents the distribution of health values ​​calculated at each time step t; R output Indicates output health; W vocab represents the mapping weight; b vocab represents the mapping bias;

[0097] (5) Based on the health value distribution and the health label, the final output health is obtained; expressed as:

[0098]

[0099] in, Indicates the final output health; Indicates the length of the time window for health status assessment; Represents the health label corresponding to the i-th sample.

[0100] Furthermore, it also includes: generating a corresponding health value curve based on the final output health; expressed as:

[0101]

[0102] in, Represents the health value curve; n represents a total of n samples in the time series; i represents the i-th sample in the time series.

[0103] Furthermore, the loss function is calculated according to the final output health and the health label; it is expressed as:

[0104]

[0105] Where L represents the mean square error loss function; b represents the number of b samples in each batch; i represents the i-th sample; Indicates the health label corresponding to the i-th sample; Represents the final output health corresponding to the i-th sample.

[0106] Furthermore, the initial health assessment model is iteratively optimized and trained according to the loss function to obtain the optimal health assessment model; specifically including:

[0107] (1) Dynamically adjust the learning rate according to the loss function; expressed as:

[0108]

[0109] in, represents the learning rate at time step t; represents the adjusted learning rate at time step t+1; factor represents the decay factor, and ;

[0110] (2) Based on the adjusted learning rate , the gradient of each model layer is calculated by the back-propagation algorithm, and the model weights are updated using the gradient descent algorithm; expressed as:

[0111]

[0112] in, represents the model weight at time step t; represents the updated model weight at time step t+1; L represents the mean square error loss function;

[0113] (3) Iteratively optimize the initial health assessment model until the training iteration reaches the preset total number of iterations, completing the training and obtaining the optimal health assessment model;

[0114] During each iterative training process, the loss function in the current round is compared with that in the previous round. If the loss difference between the two rounds is less than the preset value, it is determined that the loss function tends to be stable, and the training is terminated to obtain the optimal health assessment model.

[0115] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a health assessment method for key components of special vehicles based on orthogonal feature dynamic coupling, which has the following beneficial effects:

[0116] The model of the present invention can make full use of the vibration signal in each direction, explore the dynamic correlation and coupling relationship between the signal features in different directions, and dynamically adjust the weight of each feature according to its contribution in the current learning task, thereby promoting the flow, interaction and fusion of information within the model; specifically:

[0117] 1. The present invention proposes homogeneous and heterogeneous convolution kernels to fully extract unidirectional and orthogonal multi-scale features and coupling relationships, enhance the adaptability to multi-directional signals, and thus fully explore the dynamic correlation and coupling relationship between signal features in different directions, making health assessment more accurate.

[0118] 2. The present invention introduces a feature dynamic coupling mechanism, combines homogeneous and heterogeneous convolutions of different scales and an external knowledge base, and performs adaptive dynamic adjustment according to the importance of the features, thereby improving the generalization performance of the health assessment model.

[0119] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0120] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0121] Figure 1 A schematic flow chart of a method for health assessment of key components of special vehicles based on orthogonal feature dynamic coupling provided in an embodiment of the present invention.

[0122] Figure 2 A schematic diagram of the dynamic coupling of features combining homogeneous and heterogeneous convolution kernels provided in an embodiment of the present invention.

[0123] Figure 3 A schematic diagram of verification results of an embodiment of the present invention provided for an embodiment of the present invention.

[0124] Figure 4 The present invention provides a schematic diagram for comparing the performance of the method provided in the embodiment of the present invention with the existing method. DETAILED DESCRIPTION

[0125] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0126] The embodiment of the present invention discloses a health assessment method for key components of special vehicles based on orthogonal feature dynamic coupling, see Figure 1 As shown, the following steps are included:

[0127] S1. Obtain a vibration signal dataset of the vehicle-mounted equipment under different working conditions; the vibration signal dataset includes two orthogonal direction vibration signals of multiple samples and corresponding health labels;

[0128] S2. Batch load the vibration signal data set, and convert the samples therein into tensor form after preprocessing to obtain the initial input tensor; which serves as the input of the initial health assessment model;

[0129] S3, based on the homogeneous convolution kernel, respectively extracting features of the horizontal vibration signal and the vertical vibration signal in the initial input tensor to obtain corresponding horizontal single-direction features and vertical single-direction features;

[0130] S4, extracting features from the initial input tensor based on heterogeneous convolution kernels to obtain orthogonal direction features;

[0131] S5. By introducing external knowledge, the horizontal single-direction feature, the vertical single-direction feature and the orthogonal direction feature are coupled to obtain the final coupled feature;

[0132] S6. According to the final coupling characteristics and the health label, the final output health is obtained;

[0133] S7, calculating the loss function according to the final output health and health label;

[0134] S8. According to the loss function, the initial health assessment model is iteratively optimized and trained to obtain the optimal health assessment model;

[0135] S9. Based on the optimal health assessment model, realize the health assessment of key components of target special vehicles.

[0136] The present invention provides a health assessment method for key components of special vehicles based on dynamic coupling of orthogonal features. In order to address the problems that existing methods ignore the relationship between signal features in different directions and rely on the insufficient generalization ability of predefined models, the present invention innovatively uses homogeneous convolution kernels and heterogeneous convolution kernels to mine the dynamic correlation and coupling relationship between signal features in different directions, so that the extracted features are more comprehensive and the health assessment is more accurate. In addition, the feature dynamic coupling method is introduced to dynamically adjust the coupling weight according to the importance of the feature, which significantly improves the generalization ability of the model, thereby realizing high-precision health assessment of key components of special vehicles and providing effective support for their reliable operation.

[0137] The above-mentioned marks S1-S9 are only for the convenience of subsequent description, and do not limit the execution order of each step. Next, each of the above-mentioned steps will be described in detail.

[0138] In the above steps S1 and S2, a vibration signal data set of the vehicle-mounted equipment under different working conditions is obtained; the vibration signal data set is represented as ;in, Indicates horizontal vibration signal; Indicates vertical vibration signal; Indicates the health label corresponding to the i-th sample; Represents the total number of samples; the vibration signals in two orthogonal directions of the i-th sample are expressed as ;

[0139] To facilitate batch training, the vibration signal dataset Divide into multiple batches for loading, and the vibration signal data of each batch is represented as , where b represents the size of each batch, that is, there are b samples in each batch;

[0140] Each batch of data is processed by the model input, where R is the total number of training batches, that is, the total number of preset iterations; r represents the rth iteration. Each sample in each batch is normalized to ensure the stability of the data during training; then converted into a tensor form to obtain the initial input tensor; the initial input tensor is represented as , as the input of the initial health assessment model; each initial input tensor With shape , where L is the time step of the vibration signal.

[0141] In the above step S3, see Figure 2 As shown, including:

[0142] 1. Extract features of the horizontal vibration signal in the initial input tensor based on the homogeneous convolution kernel to obtain the corresponding horizontal unidirectional features; specifically including:

[0143] (1) The horizontal vibration signal is processed through the first-layer homogeneous convolution kernel and the corresponding weights. Convolution, normalization and activation operations are performed to obtain the horizontal first layer single-directional features, which can be expressed as:

[0144]

[0145] in, Represents the single-directional features of the first horizontal layer; represents the activation function; represents the normalization operation; Conv(·) represents the convolution process; Represents the first-layer homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0146] (2) Through the second-layer homogeneous convolution kernel and corresponding weights, the single-directional features of the first-layer After convolution, normalization and activation operations, residual connection is performed to obtain the horizontal second layer single-direction features, which can be expressed as:

[0147]

[0148] in, Represents the single-directional features of the horizontal second layer; Represents the second-layer homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0149] (3) Through the third-layer homogeneous convolution kernel and corresponding weights, the single-directional features of the second-layer After convolution, normalization and activation operations, residual connection is performed to obtain the horizontal third layer unidirectional features as the final horizontal unidirectional features; it is expressed as:

[0150]

[0151] Represents horizontal unidirectional features; Represents the third-layer homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term.

[0152] 2. Extract features of the vertical vibration signal in the initial input tensor based on the homogeneous convolution kernel to obtain the corresponding vertical unidirectional features; specifically including:

[0153] (1) The vertical vibration signal is processed through the vertical first layer homogeneous convolution kernel and corresponding weights Convolution, normalization and activation operations are performed to obtain the vertical first-layer single-direction features, which can be expressed as:

[0154]

[0155] in, Indicates the vertical first layer unidirectional characteristics; represents the activation function; represents the normalization operation; Conv(·) represents the convolution process; Represents the first vertical homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0156] (2) Through the vertical second layer homogeneous convolution kernel and corresponding weights, the single-directional features of the vertical first layer After convolution, normalization and activation function, residual connection is performed to obtain the vertical second layer single-direction features, which can be expressed as:

[0157]

[0158] in, Indicates the vertical second layer unidirectional characteristics; Represents the second vertical layer of homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0159] (3) Through the vertical third layer homogeneous convolution kernel and corresponding weights, the vertical second layer single-directional features After convolution, normalization and activation operations, residual connection is performed to obtain the vertical third layer unidirectional features as the final vertical unidirectional features; it is expressed as:

[0160]

[0161] in, Indicates vertical unidirectional features; Represents the third vertical layer of homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term.

[0162] In the above step S4, feature extraction is performed on the initial input tensor based on the heterogeneous convolution kernel to obtain orthogonal direction features; different from the homogeneous convolution kernel, the heterogeneous convolution kernel is used to establish and learn the coupling relationship of orthogonal features; the extraction process of this step specifically includes:

[0163] (1) Through the first layer of heterogeneous convolution kernels and corresponding weights, the initial input tensor is convolved, normalized and activated to obtain the first layer output features; expressed as:

[0164]

[0165] in, Represents the first layer output features; represents the activation function (ReLU); represents the normalization operation; Conv(·) represents the convolution process; Represents the first layer of heterogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0166] (2) Through the second layer of heterogeneous convolution kernels and corresponding weights, the first layer output features After convolution, normalization and activation operations, residual connection is performed to obtain the second layer output features; it is expressed as:

[0167]

[0168] in, Represents the second layer output features; Represents the second layer of heterogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0169] (3) Through the third layer of heterogeneous convolution kernels and corresponding weights, the second layer output features After convolution, normalization and activation operations, residual connection is performed to obtain the output features of the third layer, which are used as the final orthogonal direction features; it is expressed as:

[0170]

[0171] in, Represents orthogonal directional features; Represents the third layer of heterogeneous convolution kernel; express The corresponding weights; express The corresponding bias term;

[0172] The three convolution layers of the heterogeneous convolution kernel can capture the coupling relationship between the global pattern and orthogonal features in the signal through layer-by-layer calculation and feature extraction. In heterogeneous convolution, all input signals share the same set of convolution kernels, and the convolutional features are integrated across signal channels.

[0173] In the above step S5, by introducing external knowledge, the horizontal unidirectional features, the vertical unidirectional features and the orthogonal directional features are coupled to obtain the final coupled features; specifically, the process includes:

[0174] (1) The horizontal single-direction feature, the vertical single-direction feature and the orthogonal direction feature are spliced ​​to obtain a multi-scale splicing feature. Specifically, two groups of orthogonal signals are processed using independent homogeneous convolution kernels, and the convolution features of each signal are independent of each other, and multi-scale splicing is performed respectively; it is expressed as:

[0175]

[0176] Among them, F fusion input Represents multi-scale splicing features; Concat(·) represents the splicing operation; Represents horizontal unidirectional features; Indicates vertical unidirectional features; Represents orthogonal directional features;

[0177] (2) External knowledge is introduced to supplement information, and the external knowledge is encoded to obtain external knowledge encoding features; the external knowledge includes historical health data, equipment characteristics, and fault thresholds of vehicle-mounted equipment; the external knowledge encoding features are expressed as:

[0178]

[0179] K external represents the external knowledge encoding feature; Encode(·) represents the encoding operation; H represents the historical health data; P represents the equipment characteristics; T represents the fault threshold; represents the external weight matrix;

[0180] (3) Perform nonlinear mapping on the external knowledge encoding features to obtain high-dimensional features; expressed as:

[0181]

[0182] K transformer Represents high-dimensional features; represents the activation function; Represents the weight of nonlinear mapping; Represents the bias of the nonlinear mapping; and All are trainable parameters;

[0183] (4) Multi-scale splicing feature F fusion input and high-dimensional features K transformer Perform fusion processing to obtain the fusion feature F enhanced ; expressed as:

[0184]

[0185] in, and Both represent weights adaptively adjusted by feature importance, specifically:

[0186]

[0187] Among them, Attention(·) represents the self-attention operation;

[0188] (5) Fusion feature F enhanced Perform global pooling to obtain global features; expressed as:

[0189]

[0190] Among them, F global Represents global features; GlobalAvgPool(·) represents the pooling layer operation;

[0191] (6) Obtain global features F through the fully connected layer global The feature weights are , and in order to ensure the adaptability and importance of the weights, the Softmax function is used to normalize the feature weights to obtain a weight vector containing all global features; it is expressed as:

[0192]

[0193] Among them, w represents the weight vector containing all features, indicating the importance of each homogeneous and heterogeneous convolutional feature. These weights are dynamically obtained through the information extracted by the global pooling layer, and the weight of each feature will be adjusted according to its contribution in the current learning task; Linear(·) represents the fully connected layer operation;

[0194] (7) Based on the weight vector w, the multi-scale splicing feature F fusion input Perform coupling processing to obtain the initial coupling characteristics; where the weight vector w includes the horizontal unidirectional characteristics Weight , vertical single direction feature The weight w2 and orthogonal direction features Weight ,and ;

[0195] The initial coupling characteristic F combined It is expressed as:

[0196]

[0197] (8) Multi-scale splicing feature F fusion input and the initial coupling feature F combined Perform residual connection to obtain residual features; expressed as:

[0198]

[0199] Among them, F final Represents residual features; residual connections help the model better retain the original input information to enhance the training stability of the model;

[0200] (9) The residual features are transformed nonlinearly through the activation function to obtain the final coupling features; expressed as:

[0201]

[0202] Among them, F final output represents the final coupled feature.

[0203] In the above step S6, the final output health is obtained according to the final coupling feature and the health label; specifically, it includes:

[0204] (1) Based on the final coupling feature, the domain-specific unidirectional feature is decoded to obtain the unidirectional decoding feature; it is expressed as:

[0205]

[0206] Among them, R homo Represents unidirectional decoding features; represents the activation function; W homo represents the unidirectional decoding weight; b homo Indicates unidirectional decoding bias;

[0207] (2) Based on the final coupling feature, the domain-independent orthogonal direction feature is decoded to obtain the orthogonal direction decoding feature; expressed as:

[0208]

[0209] Among them, R hetero represents the orthogonal direction decoding feature; W hetero represents the decoding weight in the orthogonal direction; b hetero Indicates the decoding bias in the orthogonal direction;

[0210] (3) Decoding feature R in one direction homo And the orthogonal direction decoding feature R hetero Perform dynamic coupling to obtain the initial output health R output ; expressed as:

[0211]

[0212] in, Indicates dynamic adjustment of weights, specifically:

[0213]

[0214] Among them, Var(·) represents the variance, and the formula is as follows:

[0215]

[0216] in, express The expected value of is the mean; Represents a random variable deviation from its mean; is the square of the deviation, used to eliminate the effects of positive and negative deviations;

[0217] (4) Mapping the initial output health into a label distribution of a time series to obtain a health value distribution; expressed as:

[0218]

[0219] in, represents the distribution of health values ​​calculated at each time step t; R output Indicates output health; W vocab represents the mapping weight; b vocab represents the mapping bias;

[0220] (5) Based on the health value distribution and the health label, the final output health is obtained; expressed as:

[0221]

[0222] in, Indicates the final output health; Indicates the length of the time window for health status assessment; Represents the health label corresponding to the i-th sample.

[0223] Based on the final output health, the corresponding health value curve is generated; it is expressed as:

[0224]

[0225] in, Represents the health value curve; n represents a total of n samples in the time series; i represents the i-th sample in the time series.

[0226] In the above step S7, the loss function is calculated according to the final output health and health label; it is expressed as:

[0227]

[0228] Where L represents the mean square error loss function; b represents the number of b samples in each batch; i represents the i-th sample; Indicates the health label corresponding to the i-th sample; Represents the final output health corresponding to the i-th sample.

[0229] In the above step S8, the initial health assessment model is iteratively optimized and trained according to the loss function to obtain the optimal health assessment model; specifically, the following steps are performed:

[0230] (1) Dynamically adjust the learning rate according to the loss function; expressed as:

[0231]

[0232] in, represents the learning rate at time step t; represents the adjusted learning rate at time step t+1; factor represents the attenuation factor, which determines the rate at which the learning rate decreases, and ;

[0233] (2) Based on the adjusted learning rate , the gradient of each model layer is calculated by the back-propagation algorithm, and the model weights are updated using the gradient descent algorithm; expressed as:

[0234]

[0235] in, represents the model weight at time step t; represents the updated model weight at time step t+1; L represents the mean square error loss function; this step gradually reduces the loss function value by continuously updating the weight parameters;

[0236] (3) Iteratively optimizing and training the initial health assessment model (steps S3-S5) until the training iteration reaches a preset number of iterations, completing the training and obtaining the optimal health assessment model;

[0237] In each iterative training process, the loss function in the current round is compared with that in the previous round, that is, ,in, Represents the loss function of the current round, represents the loss function of the previous round; the question mark ? represents the preset value; if the loss difference between the two rounds is less than the preset value, it is determined that the loss function is stable, then the training can be terminated early to obtain the optimal health assessment model.

[0238] In the above step S9, based on the optimal health assessment model, a health assessment of the key components of the target special vehicle is implemented; specifically, a vibration signal data set of the target vehicle-mounted equipment is obtained, and the input tensor is obtained by the method in the above step S2, and the input tensor is input into the optimal health assessment model, and the health degree corresponding to the target vehicle-mounted equipment is output.

[0239] Next, the method provided by the present invention is described through a specific embodiment.

[0240] Taking the accelerated degradation experiment of rolling bearing as an example, the run-to-failure data collected in the experiment are used to evaluate the effectiveness of the method of the present invention. LDK UER204 bearing is used in the experimental data.

[0241] The radial force provided by the hydraulic loading system of the accelerated degradation experiment is applied to the outer shell of the tested bearing, perpendicular to the direction of gravity. The vibration signal of the tested bearing is collected by vertical and horizontal accelerometers with a sampling frequency of 25.6 kHz, that is, 2560 data points are recorded every 0.1 second, collected once every 60 seconds, and each collection time is 1.28 seconds. Three different working conditions were simulated by setting different speeds and different radial forces. The speeds were set to 2100rpm, 2250rpm, and 2400rpm, respectively, and the radial forces were set to 12KN, 11KN, and 10KN, respectively, and 15 bearings were tested.

[0242] The experiment compares the method of the present invention with the existing learning methods. Method A constructs a model structure similar to the method of the present invention, using homogeneous and heterogeneous convolution kernels, but does not use feature dynamic coupling coding, only using average weighted summation. Method B constructs a model structure with two groups of ordinary convolution kernels in each of the three layers, and uses feature dynamic coupling coding. Method C constructs a model structure with two groups of ordinary convolution kernels in each of the three layers, using average weighted summation. Method D uses one group of ordinary convolution kernels in each of the three layers for learning.

[0243] In the experimental results, the present invention uses the root mean square error (RMSE) and the symmetric mean absolute percentage error (sMAPE) to evaluate the performance of the proposed model, and the calculation formula is:

[0244]

[0245]

[0246] in, Indicates the number of samples, which here indicates the total number of samples in the test set; represents the predicted remaining useful life value of the i-th sample; Represents the actual remaining useful life value of the i-th sample.

[0247] The experimental conditions are shown in Table 1, the constructed orthogonal characteristic dynamic coupling model learning scenario is shown in Table 2, and the model parameter configuration is shown in Table 3. The experimental results are summarized in Table 4. Figure 3 and Figure 4 .

[0248] Table 1. Bearing working condition configuration in the embodiment

[0249] Simulation conditions Speed Radial force Bearing No. Condition 1 2100rpm 12KN Bearing 1-1~Bearing 1-5 Condition 2 2250rpm 11KN Bearing 2-1~Bearing 2-5 Condition 3 2400rpm 10KN Bearing 3-1~Bearing 3-5

[0250] Table 2. Learning scenarios constructed in the embodiment

[0251] Training Dataset Test Dataset Bearing 1-1, bearing 2-1, bearing 3-2 Bearing 1-3, bearing 2-2, bearing 3-3 Bearing 1-4, bearing 2-4, bearing 3-4 Bearing 1-5, bearing 2-5, bearing 3-5 Bearing 1-2 Bearing 2-3 Bearing 3-1

[0252] Table 3. Model parameter configuration

[0253] Model Layer Parameter settings Input Layer - Heterogeneous Convolutional Layer #1 Number of convolution kernels: 64 Convolution kernel size: 3, 5, 7 Activation function: ReLu Homogeneous convolution kernel #1 Number of convolution kernels: 64 Convolution kernel size: 3 Attention mechanism: SE module Activation function: ReLu Batch Normalization Layer #1 - Max Pooling Layer #1 Pooling size: 2 Heterogeneous Convolutional Layer #2 Number of convolution kernels: 128 Convolution kernel size: 3, 5, 7 Activation function: ReLu Homogeneous convolution kernel #2 Number of convolution kernels: 128 Convolution kernel size: 3 Attention mechanism: SE module Activation function: ReLu Batch Normalization Layer #2 - Max Pooling Layer #2 Pooling size: 2 Heterogeneous Convolutional Layer #3 Number of convolution kernels: 256 Convolution kernel size: 3, 5, 7 Activation function: ReLu Homogeneous convolution kernel #3 Number of convolution kernels: 256 Convolution kernel size: 3 Attention mechanism: SE module Activation function: ReLu Batch Normalization Layer #3 - Global Average Pooling - Fully connected layer #1 Number of neurons: 256 Activation function: ReLu Fully connected layer #2 Number of neurons: 1 Activation function: -

[0254] Table 4. Evaluation results of the performance of the method of the present invention

[0255] method Root Mean Square Error (RMSE) Mean absolute percentage error (sMAPE) This method 102.945 ± 77.0974 54.5902 ± 18.9962 Method A 167.85545 ± 128.62715 68.06595 ± 44.00185 Method B 324.3999 ± 264.0986 75.3196 ± 43.3941 Method C 437.79115 ± 365.47245 109.1962 ± 50.363 Method D 304.6191 ± 260.4265 107.00495 ± 63.88215

[0256] It can be seen from the experimental results that the present invention has better health assessment performance in the learning process, and the method of the present invention has obvious advantages over other methods.

[0257] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0258] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one 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 present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A health assessment method for key components of special vehicles based on orthogonal feature dynamic coupling, characterized in that: The following steps are involved: Acquire a vibration signal data set of the vehicle-mounted equipment under different working conditions; the vibration signal data set includes two orthogonal direction vibration signals of multiple samples and corresponding health labels; Batch loading the vibration signal data set, and preprocessing the samples therein and converting them into a tensor form to obtain an initial input tensor; As input to initial models for health assessment; Based on the homogeneous convolution kernel, the horizontal vibration signal and the vertical vibration signal in the initial input tensor are respectively extracted to obtain the corresponding horizontal single-direction feature and vertical single-direction feature; Performing feature extraction on the initial input tensor based on a heterogeneous convolution kernel to obtain orthogonal direction features; By introducing external knowledge, coupling processing is performed on the horizontal unidirectional feature, the vertical unidirectional feature and the orthogonal directional feature to obtain a final coupling feature; According to the final coupling feature, combined with the health label, a final output health is obtained; Calculate a loss function according to the final output health and the health label; According to the loss function, the initial health assessment model is iteratively optimized and trained to obtain an optimal health assessment model; Based on the optimal health assessment model, health assessment of key components of target special vehicles is achieved.

2. The method for health assessment of key components of special vehicles based on orthogonal feature dynamic coupling according to claim 1 is characterized in that: The vibration signal data set is batch loaded, and two orthogonal vibration signals are preprocessed to obtain an initial input tensor; Specifically include: The vibration signal dataset is represented as ;in, Indicates horizontal vibration signal; Indicates vertical vibration signal; Indicates the health label corresponding to the i-th sample; Represents the total number of samples; the vibration signals in two orthogonal directions of the i-th sample are expressed as ; The vibration signal dataset Divide into multiple batches for loading, and the vibration signal data of each batch is represented as , where b means there are b samples in each batch; Each sample in each batch is normalized and converted into a tensor form to obtain the initial input tensor; the initial input tensor is expressed as .

3. The health assessment method for key components of special vehicles based on orthogonal feature dynamic coupling according to claim 1 is characterized in that: Based on the homogeneous convolution kernel, the horizontal vibration signal in the initial input tensor is extracted to obtain the corresponding horizontal single-direction feature; Specifically include: (1) The horizontal vibration signal is processed through the first-layer homogeneous convolution kernel and the corresponding weights. Convolution, normalization and activation operations are performed to obtain the horizontal first layer single-directional features, which can be expressed as: ； in, Represents the single-directional features of the first horizontal layer; represents the activation function; represents the normalization operation; Conv(·) represents the convolution process; Represents the first-layer homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term; (2) Through the second-level homogeneous convolution kernel and corresponding weights, the single-directional features of the first-level After convolution, normalization and activation operations, residual connection is performed to obtain the horizontal second layer single-direction features, which can be expressed as: ； in, Represents the single-directional features of the horizontal second layer; Represents the second-layer homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term; (3) Through the third-layer homogeneous convolution kernel and corresponding weights, the single-directional features of the second-layer After convolution, normalization and activation operations, residual connection is performed to obtain the horizontal third layer unidirectional features as the final horizontal unidirectional features; it is expressed as: ； Represents horizontal unidirectional features; Represents the third-layer homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term.

4. The health assessment method for key components of special vehicles based on orthogonal feature dynamic coupling according to claim 1 is characterized in that: Based on the homogeneous convolution kernel, the vertical vibration signal in the initial input tensor is extracted to obtain the corresponding vertical unidirectional features; specifically, it includes: (1) The vertical vibration signal is processed through the vertical first layer homogeneous convolution kernel and corresponding weights Convolution, normalization and activation operations are performed to obtain the vertical first-layer single-direction features, which can be expressed as: ； in, Indicates the vertical first layer unidirectional characteristics; represents the activation function; represents the normalization operation; Conv(·) represents the convolution process; Represents the first vertical homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term; (2) Through the vertical second layer homogeneous convolution kernel and corresponding weights, the single-directional features of the vertical first layer After convolution, normalization and activation function, residual connection is performed to obtain the second vertical layer single-direction features, which can be expressed as: ； in, Indicates the vertical second-layer unidirectional characteristics; Represents the second vertical layer of homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term; (3) Through the vertical third layer homogeneous convolution kernel and corresponding weights, the vertical second layer single-directional features After convolution, normalization and activation operations, residual connection is performed to obtain the vertical third layer unidirectional features as the final vertical unidirectional features; it is expressed as: ； in, Indicates vertical unidirectional features; Represents the third vertical layer of homogeneous convolution kernel; express The corresponding weights; express The corresponding bias term.

5. The special vehicle key component health assessment method based on orthogonal feature dynamic coupling according to claim 1 is characterized in that: The method of extracting features from the initial input tensor based on the heterogeneous convolution kernel to obtain orthogonal direction features specifically includes: (1) Through the first layer of heterogeneous convolution kernels and corresponding weights, the initial input tensor is convolved, normalized and activated to obtain the first layer output features; expressed as: ； in, Represents the first layer output features; represents the activation function; represents the normalization operation; Conv(·) represents the convolution process; Represents the first layer of heterogeneous convolution kernel; express The corresponding weights; express The corresponding bias term; represents the initial input tensor; (2) Through the second layer of heterogeneous convolution kernels and corresponding weights, the first layer output features After convolution, normalization and activation operations, residual connection is performed to obtain the second layer output features; it is expressed as: ； in, Represents the second layer output features; Represents the second layer of heterogeneous convolution kernel; express The corresponding weights; express The corresponding bias term; (3) Through the third layer of heterogeneous convolution kernels and corresponding weights, the second layer output features After convolution, normalization and activation operations, residual connection is performed to obtain the output features of the third layer, which are used as the final orthogonal direction features; it is expressed as: ； in, Represents orthogonal directional features; Represents the third layer of heterogeneous convolution kernel; express The corresponding weights; express The corresponding bias term.

6. The method for health assessment of key components of special vehicles based on orthogonal feature dynamic coupling according to claim 1 is characterized in that: The method of introducing external knowledge to couple the horizontal unidirectional feature, the vertical unidirectional feature and the orthogonal directional feature to obtain the final coupled feature specifically includes: (1) The horizontal single-direction feature, the vertical single-direction feature, and the orthogonal direction feature are spliced ​​to obtain a multi-scale splicing feature; expressed as: ； Among them, F fusion input Represents multi-scale splicing features; Concat(·) represents the splicing operation; Represents horizontal unidirectional features; Indicates vertical unidirectional features; Represents orthogonal directional features; (2) Encoding the external knowledge to obtain external knowledge encoding features; the external knowledge includes historical health data, equipment characteristics, and fault thresholds of the vehicle-mounted equipment; the external knowledge encoding features are expressed as: ； K external represents the external knowledge encoding feature; Encode(·) represents the encoding operation; H represents the historical health data; P represents the equipment characteristics; T represents the fault threshold; represents the external weight matrix; (3) Perform nonlinear mapping on the external knowledge encoding features to obtain high-dimensional features; expressed as: ； K transformer Represents high-dimensional features; represents the activation function; Represents the weight of nonlinear mapping; Represents the bias of the nonlinear mapping; (4) Multi-scale splicing feature F fusion input and high-dimensional features K transformer Perform fusion processing to obtain the fusion feature F enhanced ; expressed as: ； in, and Both represent weights adaptively adjusted by feature importance, specifically: ； Among them, Attention(·) represents the self-attention operation; (5) Fusion feature F enhanced Perform global pooling to obtain global features; expressed as: ； Among them, F global Represents global features; GlobalAvgPool(·) represents the pooling layer operation; (6) Obtain global features F through the fully connected layer global The feature weights are normalized using the Softmax function to obtain a weight vector containing all global features; it is expressed as: ； Where w represents the weight vector containing all features; Linear(·) represents the fully connected layer operation; (7) Based on the weight vector w, the multi-scale splicing feature F fusion input Perform coupling processing to obtain the initial coupling characteristics; where the weight vector w includes the horizontal unidirectional characteristics Weight , vertical single direction feature The weight w2 and orthogonal direction features Weight ,and ; The initial coupling characteristic F combined It is expressed as: ； (8) Multi-scale splicing feature F fusion input and the initial coupling feature F combined Perform residual connection to obtain residual features; expressed as: ； Among them, F final represents the residual feature; (9) The residual features are transformed nonlinearly through the activation function to obtain the final coupling features; expressed as: ； Among them, F final output represents the final coupled feature.

7. The method for health assessment of key components of special vehicles based on orthogonal feature dynamic coupling according to claim 1 is characterized in that: The step of obtaining the final output health degree according to the final coupling feature and the health degree label specifically includes: (1) Based on the final coupling feature, the unidirectional feature is decoded to obtain the unidirectional decoding feature; it is expressed as: ； Among them, R homo Represents unidirectional decoding features; represents the activation function; W homo represents the unidirectional decoding weight; b homo Indicates unidirectional decoding bias; F final output represents the final coupling characteristics; (2) Based on the final coupling feature, the orthogonal direction feature is decoded to obtain the orthogonal direction decoding feature; which is expressed as: ； Among them, R hetero represents the orthogonal direction decoding feature; W hetero represents the decoding weight in the orthogonal direction; b hetero Indicates the decoding bias in the orthogonal direction; (3) Decoding feature R in one direction homo And the orthogonal direction decoding feature R hetero Perform dynamic coupling to obtain the initial output health R output ; expressed as: ； in, Indicates dynamic adjustment of weights, specifically: ； Where Var(·) represents variance; (4) Mapping the initial output health into a label distribution of a time series to obtain a health value distribution; expressed as: ； in, represents the distribution of health values ​​calculated at each time step t; R output Indicates output health; W vocab represents the mapping weight; b vocab represents the mapping bias; (5) Based on the health value distribution and the health label, the final output health is obtained; expressed as: ； in, Indicates the final output health; Indicates the length of the time window for health status assessment; Represents the health label corresponding to the i-th sample.

8. The method for health assessment of key components of special vehicles based on orthogonal feature dynamic coupling according to claim 7 is characterized in that: Also includes: Based on the final output health, a corresponding health value curve is generated; it is expressed as: ； in, Represents the health value curve; n represents a total of n samples in the time series; i represents the i-th sample in the time series.

9. The method for health assessment of key components of special vehicles based on orthogonal feature dynamic coupling according to claim 1 is characterized in that: The loss function is calculated according to the final output health and the health label; it is expressed as: ； Where L represents the mean square error loss function; b represents the number of b samples in each batch; i represents the i-th sample; Indicates the health label corresponding to the i-th sample; Represents the final output health corresponding to the i-th sample.

10. The method for health assessment of key components of special vehicles based on orthogonal feature dynamic coupling according to claim 1 is characterized in that: The iterative optimization training of the initial health assessment model according to the loss function to obtain the optimal health assessment model specifically includes: (1) Dynamically adjust the learning rate according to the loss function; expressed as: ； in, represents the learning rate at time step t; represents the adjusted learning rate at time step t+1; factor represents the decay factor, and ; (2) Based on the adjusted learning rate , the gradient of each model layer is calculated by the back-propagation algorithm, and the model weights are updated using the gradient descent algorithm; expressed as: ； in, represents the model weight at time step t; represents the updated model weight at time step t+1; L represents the mean square error loss function; (3) Iteratively optimize the initial health assessment model until the training iteration reaches the preset total number of iterations, completing the training and obtaining the optimal health assessment model; During each iterative training process, the loss function in the current round is compared with that in the previous round. If the loss difference between the two rounds is less than the preset value, it is determined that the loss function tends to be stable, and the training is terminated to obtain the optimal health assessment model.

Images