A decoupling recognition method for continuous operating conditions of diesel engines based on graph self-attention network

Abstract

A decoupling recognition method for continuous operating conditions of diesel engines based on graph self-attention networks belongs to the technical field of equipment condition monitoring and diagnosis. First, experimental operating conditions and condition labels are designed, and the training set and test set are divided according to the condition distribution. Secondly, a self-attention network layer is established to calculate the similarity between angular domain signal segments. Subsequently, a graph convolutional network layer is established to aggregate neighbor node information. Then, a decoupled condition space is established to map the signal features to the decoupled condition space through dimensionality reduction. Finally, a condition recognition network model that ingeniously combines the self-attention mechanism and the graph convolutional algorithm is formed, trained, and saved. When a new sample is input into the saved condition recognition model, the specific rotational speed and load values of the current signal will be given, realizing the decoupling and accurate recognition of multiple condition parameters.

Description

Technical Field

[0001] The present invention relates to a method for identifying working conditions, in particular to a method for constructing a continuous working condition decoupling identification model, which is applicable to the technical field of equipment condition monitoring and diagnosis. Background Art

[0002] As one of the core power equipment in the large-scale industrial field, diesel engines have a very wide range of applications. Conducting condition monitoring and fault diagnosis on them can timely and correctly diagnose abnormal or fault states, improve the reliability and safety of equipment operation, and obtain greater economic benefits. Since the vibration signals of diesel engine equipment belong to typical non-stationary vibration signals, which are composed of multiple impacts related to the period and have complex impact characteristics, intelligent fault diagnosis systems often cannot distinguish between working condition characteristics and fault characteristics. For example, an increase in the degree of fault and an increase in working condition conditions will both lead to an increase in the peak value of the vibration signal, and the causal relationship between the two is coupled. Without working condition information as a reference, it is difficult to correctly diagnose the health status of the diesel engine unit. Therefore, the accurate identification of diesel engine working conditions is of great significance.

[0003] The method for identifying working conditions based on the knowledge fusion deep learning network has been proven to be able to well extract the deep features in the signal, enabling the expression of working condition signals, becoming a research hotspot in the field of condition monitoring and diagnosis, and relatively rich technical achievements have been formed. However, most working condition identification models use power as the identification parameter, divide the power parameter into several discrete intervals, and obtain the power interval where the current state is located through a classification model. Describing the working condition only by power will cause the coupling of rotational speed and load parameters, and the working condition parameters should be continuous variables. Therefore, in the monitoring and diagnosis technologies related to mechanical equipment, there is still a lack of a feasible method as described above for constructing a continuous working condition decoupling identification model that can identify continuous working conditions and decouple specific rotational speed and load parameters. As an algorithm for Laplacian eigenmaps, the graph convolutional network models on graph data, constructs the relationship between data from a local perspective, makes the points with relationships as close as possible in the mapped space, and realizes a better manifold learning mapping relationship. At the same time, the self-attention mechanism is a powerful tool for capturing the internal correlation of data features. Therefore, combining the self-attention mechanism and the graph convolutional network method to enhance the representation of the working condition features of the signal becomes a feasible solution.

[0004] The present invention proposes a continuous working condition model decoupled from rotational speed and load parameters, and constructs a working condition decoupling model based on the graph self-attention network. This method extracts the similarity of signals in different time intervals through the self-attention mechanism, uses the self-attention coefficient as the adjacency matrix, enhances the feature representation of samples through the graph convolutional algorithm, and finally maps the sample features to a working condition decoupling space to achieve the decoupling and accurate identification of rotational speed and load. Summary of the Invention

[0005] The object of the present invention is to provide an accurate and effective decoupling recognition method for multi-parameter working condition recognition of mechanical equipment vibration signals, and to complete the recognition of continuous values.

[0006] The object of the present invention is achieved through the following technical solutions:

[0007] First, design experimental working conditions to collect experimental data on the continuous working condition changes of a diesel engine, and divide the training set samples and test set samples. Moreover, divide the cylinder head vibration signal into angular domain periodic signals, and perform noise reduction filtering and normalization processing on them to make them conform to the input form of the neural network model.

[0008] Secondly, use two independent parameters to represent the sample working condition label: γ = [α, β], and perform normalization processing with the rated speed and rated load respectively. That is, without considering overspeed and overload operation, the value ranges of α and β are both [0, 100].

[0009] Thirdly, establish a self-attention mechanism network model, divide the sample signal into m angular domain signal segments, multiply them with the weight matrix to obtain Queries vectors and Keys vectors, and then calculate the similarity between the Queries vectors and Keys vectors to obtain the self-attention coefficient matrix.

[0010] Fourthly, establish a graph convolutional network model, use the angular domain signal segments as nodes, and the self-attention coefficient matrix as the adjacency matrix. Through the graph convolutional network algorithm, aggregate features with higher similarity to itself and greater contribution to classification, and finally fuse the node signals, reorganize the angular domain signals, and enhance the feature representation of the signals.

[0011] Finally, establish a working condition decoupling space, map the signal features to the working condition decoupling space for dimensionality reduction, and complete the decoupling and recognition of the working conditions.

[0012] A continuous working condition decoupling recognition model method combining a self-attention mechanism and a graph convolutional network algorithm, characterized by including the following steps:

[0013] The first step: Establish a sample data set

[0014] 1.1 Collect experimental data on the continuous working condition changes of a diesel engine according to the working condition distribution, and organize to obtain the total sample set D:

[0015] D = {F 1 , F 2 , …, F t , …, F T} (1)

[0016]

[0017]

[0018] Among them, D is the total sample set, and F t represents the sample set of the t-th type of working condition parameters, T is the number of working condition categories, represents the i-th sample under the t-th type of working condition parameters, and n represents the number of samples under the t-th type of working condition parameters, represents the j-th sequence point of the i-th sample under the t-th type of working condition parameters, represents the K-th sequence point of the i-th sample under the t-th type of working condition parameters, and K is the number of sequence points included in each sample, which is numerically equal to the number of sequence points collected by the sensor during a complete working cycle of the machine operation.

[0019] 1.2 Normalize the signal after filtering and noise reduction. Denote x as a general sample, that is, update each sample x in the following way:

[0020] f(x): x → x new (4)

[0021]

[0022] Among them, x new is the normalized sample, x new ∈[-1, 1], x min is the minimum value in the sample x sequence, and x max is the maximum value in the sample x sequence.

[0023] 1.3 Divide the total sample set D into a training set and a test set according to the sample categories, with a ratio of 3:1, and the working condition categories corresponding to the training set samples are different from those of the test set. For example, there are a total of 24 working condition categories. Divide the samples of 18 working conditions into the training set, and the samples of the remaining 6 working conditions into the test set.

[0024] Second step: Design the sample working condition label

[0025] Since the dimensions of the working condition parameters are different, it will lead to different precisions of the output errors of different channels of the network. Use two independent parameters to represent the sample working condition label: γ = [α, β], and perform normalization processing using the rated speed and the rated load respectively. The calculation process is as follows:

[0026]

[0027]

[0028] Among them, V is the current speed of the signal, V idling is the idle speed of the diesel engine, and V rated is the rated speed of the diesel engine; L is the current load of the signal, and L min is the minimum load of the diesel engine. Generally, Lmin = 0, L rated is the rated load of the diesel engine, L rated The calculation process is as follows:

[0029]

[0030] Among them, P rated is the rated power of the diesel engine.

[0031] Step 3: Establish a self-attention mechanism network model

[0032] There are internal correlations between the hidden features of different angular domain signal segments. For the same classification task, the hidden features contain different classification contribution degrees and have different effects on the recognition results. The self-attention mechanism is adopted to mine the key information in the signals of each time interval.

[0033] At the input layer, the sample x is divided into m angular domain signal segments with a dimension of d, and the update process is as follows:

[0034] x = [x 1 , x 2 , …, x i , …, x m T ∈R m×d (9)

[0035] Among them, x i is the i-th segment of the sample x.

[0036] Calculate the Queries vector and the Keys vector as follows:

[0037]

[0038]

[0039] Among them, W q and are linear transformation matrices with a dimension of d×d k , the dimensions of the Queries vector and the Keys vector are m×d k , where the Queries vector represents the feature expression of the current signal segment, the Keys vector represents the feature expressions of other signal segments in the input, calculate the similarity between the Queries vector and all Keys vectors, and then apply an activation function to obtain the self-attention relationship matrix Att of the signal segment:

[0040]

[0041] Among them, Softmax(·) is the softmax activation function, d k ​is the dimension of the Queries vector and the Keys vector.

[0042] Step 4: Establish a graph convolutional network model

[0043] Establish the graph data G, and the establishment process is as follows:

[0044] G = (x, Att) (13)

[0045] Among them, the angular domain signal segment x is used as the node, with a total of m nodes, and the self-attention coefficient Att is used as the edge. The nodes in the graph aggregate the features of neighbor nodes through the graph convolution algorithm to generate new node representations. The update process of the graph convolution algorithm is as follows:

[0046] h = σ(L sym xW h ) (14)

[0047]

[0048] Among them, h is the output matrix of the graph convolutional layer, σ(·) is the non-linear activation function, L sym is the Symmetric normalized Laplacian matrix, W h is the trainable weight matrix, and its dimension is d×d. is the adjacency matrix after adding self-loops. I is the identity matrix, Dig = diag(d(v 1 ), d(v 2 ), …, d(v n )) ∈ R n×n is the degree matrix of G, where d(v 1 ) = ∑ j∈V a ij is the degree of node v 1 . is 's degree matrix. The feature matrix x completes the process of propagating and receiving features between neighbor nodes by left-multiplying the adjacency matrix, and then right-multiplies a trainable weight matrix W h , completes the linear transformation, and finally is activated by the non-linear activation function to complete the feature aggregation.

[0049] Finally, arrange and combine each signal segment in the original order to obtain the signal enhancement matrix D':

[0050] X' = Reshape(h) ∈ R 1×K (16)

[0051]

[0052] D' = {F' 1, F′ 2 , …, F′ T} (18)

[0053] Among them, Reshape() is a matrix dimension transformation function, X′ is the output vector after the dimension transformation of a sample, and F′ t is the sample data set of the t-th working condition after signal enhancement, where is the first sample under the t-th working condition after signal enhancement, n represents the number of samples under the t-th working condition parameters, K represents the number of sequence points included in each sample, which is numerically equal to the number of sequence points collected by the sensor during a complete working cycle of the machine operation, D′ is the total sample data set after signal enhancement, and T is the number of working condition categories.

[0054] Step 5: Establish a working condition decoupling space

[0055] Define the network model classifier as the working condition decoupling space, which consists of a single fully connected layer. The number of its nodes is numerically equal to 2, namely the rotational speed and load channels respectively, and the activation function is selected as ReLu. The working condition decoupling space quantifies the working condition labels. Taking the rotational speed and load channels as the reference axes and the specific parameter values as the scales, the sample data is dimensionally reduced and mapped into the working condition decoupling space to visualize the numerical values of sample working condition recognition, and complete the decoupling and continuous recognition of the rotational speed and load parameters.

[0056] Step 6: Establish a final working condition decoupling recognition model.

[0057] The final recognition model is established as a two-channel model. Each channel consists of a self-attention network layer, a graph convolutional network layer, and a working condition decoupling space classifier. The input of the final model is the sample after data preprocessing in the above steps 1.1, 1.2, and 1.3, and the output is the label of the sample in the second step above. The model objective is to minimize the mean absolute error mae of the predicted value. The loss function loss selects mae, and the optimization method selects Adam. Finally, train the working condition recognition model and stop training after the mae gradient descent reaches 0, and save the model.

[0058] Step 7: Decouple and recognize working condition parameters.

[0059] When the machine is in the subsequent working process, a signal of a working cycle can be arbitrarily extracted as a sample. After normalization in step 1.2, it is input into the model saved in step 6, and the rotational speed and load parameter values of the sample are automatically displayed in the working condition decoupling space, and these values are continuous values.

[0060] The present invention proposes a method for constructing a continuous working condition decoupling recognition model based on a graph self-attention network. When a new sample is input into the saved working condition recognition model, it will give the specific rotational speed and load values of the current signal, realizing the decoupling and accurate recognition of multiple working condition parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following-described drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0062] Figure 1 is a flowchart of a method for constructing a continuous operating condition decoupling recognition model based on a graph self-attention network provided according to an example of the present application;

[0063] Figure 2 is a schematic diagram of experimental operating conditions and label design and a schematic diagram of dataset division provided according to an example of the present application;

[0064] Figure 3 is a structural diagram of a self-attention network layer provided according to an example of the present application;

[0065] Figure 4 is a structural diagram of a graph convolutional network layer provided according to an example of the present application;

[0066] Figure 5 is a schematic diagram of the structure of a continuous operating condition decoupling recognition model based on a graph self-attention network provided according to an example of the present application;

[0067] Figure 6 is a schematic diagram of the projection of the model recognition result - operating condition decoupling space provided according to an example of the present application;

[0068] Figure 7 is a schematic diagram of the application of the recognition model provided according to an example of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0069] To better understand the technical solutions of the present invention, taking the vibration signal of the cylinder head of a TBD234V12 diesel engine as a specific implementation object, a method for constructing a continuous operating condition decoupling recognition model proposed by the present invention is calculated and implemented.

[0070] Figure 1 is a flowchart of a method for constructing a continuous operating condition decoupling recognition model based on a graph self-attention network provided by the present application. See Figure 1 , the implementation process and results of the present invention are as follows:

[0071] First step: Establish a sample dataset

[0072] Arrange vibration acceleration sensors at the cylinder head position of the engine, collect cylinder vibration signals, with a sampling frequency of 51200 Hz, and divide them into angular domain signals with a crankshaft angle of 720° as a cycle.

[0073] 1.1 Collect the experimental data of the continuous operating conditions of the diesel engine according to the distribution of operating conditions, and organize to obtain the total sample set D:

[0074] D = {F 1 , F 2 , …, F t , …, F T} (1)

[0075]

[0076]

[0077] Among them, D is the total sample set, F t represents the sample set of the t-th type of operating condition parameters, T is the number of operating condition categories, represents the i-th sample under the t-th type of operating condition parameters, n represents the number of samples under the t-th type of operating condition parameters, represents the j-th sequence point of the i-th sample under the t-th type of operating condition parameters, represents the K-th sequence point of the i-th sample under the t-th type of operating condition parameters, K is the number of sequence points included in each sample, and numerically equals the number of sequence points collected by the sensor during a complete working cycle of the machine operation.

[0078] In this example, the rotational speed and load are set as Figure 2 shown in a, where there are a total of 24 continuously changing operating conditions, which are the rotational speed change data under a fixed load and the load change data under a fixed rotational speed, n = 100, K = 4032.

[0079] 1.2 Normalize the signal after filtering and noise reduction processing. Denote x as a general sample, that is, update each sample x in the following way:

[0080] f(x): x → x new (4)

[0081]

[0082] Among them, x new is the normalized sample, x new ∈[-1, 1], x min is the minimum value in the sample x sequence, and x max is the maximum value in the sample x sequence.

[0083] 1.3 Divide the total sample set D into a training set and a test set according to the sample categories, with a ratio of 3:1, and the operating condition categories corresponding to the training set samples are different from those of the test set. For example, there are a total of 24 operating condition categories. Divide the samples of 18 operating conditions into the training set, and the samples of the remaining 6 operating conditions into the test set.

[0084] The dataset division of this example is as Figure 2 shown in (b).

[0085] Step 2: Design the sample working condition labels

[0086] Since the dimensions of the working condition parameters are different, it will lead to different precisions of the output errors of different channels of the network. Under the test of the diesel engine experimental data in this example, the mae of the speed channel output is ±10, and the difference between the upper and lower limits of the speed is 800 rpm, so the relative error of the speed is ±1.25%, while the mae of the load channel output is ±10, and the difference between the upper and lower limits of the load is 2375 N·m, and the relative error of the load is 0.42%. Therefore, two independent parameters are used to represent the sample working condition labels: γ = [α, β], and they are normalized by the rated speed and the rated load respectively. The calculation process is as follows:

[0087]

[0088]

[0089] Among them, V is the current speed of the signal, V idling is the idling speed of the diesel engine, V rated is the rated speed of the diesel engine; L is the current load of the signal, L min is the minimum load of the diesel engine. Generally, L min = 0, L rated is the rated load of the diesel engine, L rated The calculation process is as follows:

[0090]

[0091] Among them, P rated is the rated power of the diesel engine.

[0092] In this example, V idling = 700 rpm, V rated = 1500 rpm, P rated = 373 kw, L min = 0 N·m, L rated = 2375 N·m. The label design is as Figure 2 shown in (c).

[0093] Step 3: Establish a self-attention mechanism network model

[0094] There is an internal correlation between the hidden features of different angular domain signal segments. For the same classification task, the hidden features contain different classification contribution degrees and have different influences on the recognition results. The self-attention mechanism is adopted to mine the key information in the signals of each time interval.

[0095] At the input layer, the sample x is divided into m angular domain signal segments with dimension d, and the update process is as follows:

[0096] x = [x 1 , x 2 , …, x i , …, x m T ∈R m×d (9)

[0097] where x i is the i-th segment of the sample x.

[0098] Calculate the Queries vector and the Keys vector as follows:

[0099]

[0100]

[0101] where W q and are linear transformation matrices with dimension d×d k , the dimensions of the Queries vector and the Keys vector are m×d k , where the Queries vector represents the feature expression of the current signal segment, the Keys vector represents the feature expressions of other signal segments in the input, calculate the similarity between the Queries vector and all Keys vectors, and then apply an activation function to obtain the self-attention relationship matrix Att of the signal segment:

[0102]

[0103] where Softmax(·) is the softmax activation function, d k is the dimension of the Queries vector and the Keys vector.

[0104] The calculation process of the self-attention layer in this example is as Figure 3 shown.

[0105] Step 4: Build a graph convolutional network model

[0106] Build the graph data G, and the building process is as follows:

[0107] G = (x, Att) (13)

[0108] where the angular domain signal segments x are used as nodes, with a total of m nodes, and the self-attention coefficients Att are used as edges. The nodes in the graph aggregate the features of neighbor nodes through the graph convolutional algorithm to generate new node representations. The update process of the graph convolutional algorithm is as follows:

[0109] h = σ(L​sym xW h ) (14)

[0110]

[0111] where h is the output matrix of the graph convolutional layer, σ(·) is the non-linear activation function, L sym is the Symmetric normalized Laplacian matrix, W h is the trainable weight matrix with dimension d×d, is the adjacency matrix after adding self-loops, I is the identity matrix, Dig = diag(d(v 1 ), d(v 2 ), …, d(v n )) ∈ R n×n is the degree matrix of G, where d(v 1 ) = ∑ j∈V a ij is the degree of node v 1 , is 's degree matrix. The feature matrix x completes the process of propagating and receiving features between neighbor nodes by left-multiplying the adjacency matrix, and then right-multiplies a trainable weight matrix W h to complete the linear transformation, and finally is activated by the non-linear activation function to complete feature aggregation.

[0112] Finally, the individual signal segments are arranged and combined in the original order to obtain the signal enhancement matrix D':

[0113] X' = Reshape(h) ∈ R 1×K (16)

[0114]

[0115] D' = {F' 1 , F' 2 , …, F' T} (18)

[0116] where Reshape() is the matrix dimension transformation function, X' is the output vector after dimension transformation of a sample, F' t is the dataset of the t-th type of working condition samples after signal enhancement, where is the first sample under the t-th type of working condition after signal enhancement, n represents the number of samples under the t-th type of working condition parameters, K represents the number of sequence points included in each sample, which is numerically equal to the number of sequence points collected by the sensor during a complete working cycle of the machine operation, D' is the total sample dataset after signal enhancement, and T is the number of working condition categories.

[0117] The calculation process of the graph convolutional network layer in this example is as follows Figure 4 shown.

[0118] Step 5: Establish the working condition decoupling space

[0119] Define the network model classifier as the working condition decoupling space, which is composed of a fully connected layer. The number of its nodes is numerically equal to 2, namely the rotational speed and load channels respectively. The activation function is selected as ReLu. The working condition decoupling space quantifies the working condition labels. Taking the rotational speed and load channels as the reference axes and the specific parameter values as the scales, the sample data is dimensionally reduced and mapped into the working condition decoupling space, visualizing the numerical values of the sample working condition recognition, and completing the decoupling and continuous recognition of the rotational speed and load parameters.

[0120] Step 6: Establish the final working condition decoupling recognition model.

[0121] The final recognition model is established as a two-channel model. Each channel is composed of a self-attention network layer, a graph convolutional network layer and a working condition decoupling space classifier. The input of the final model is the sample after data preprocessing in the above steps 1.1, 1.2 and 1.3, and the output is the label of the sample in the second step above. The goal of the model is to minimize the mean absolute error mae of the predicted value. The loss function loss selects the general mae, and the optimization method selects the general Adam. When the final diagnosis model is in training, stop training after the mae gradient descent reaches 0, and save the model.

[0122] The model structure established in this example is as follows Figure 5 shown. The visualization effect of the projection of the dataset in the working condition decoupling space is as follows Figure 6 shown.

[0123] Step 7: Decoupling and recognition of working condition parameters

[0124] When the machine is in the subsequent working process, a signal of an arbitrary working cycle can be extracted as a sample. After normalization in 2.2, it is input into the model saved in the sixth step, and the rotational speed and load parameter values of the sample are automatically displayed in the working condition decoupling space, and these values are continuous values.

[0125] The application example of this example is as follows Figure 7 shown. The actual working condition is [1200 rpm, 800 N·m], the actual label is [62.5, 33.68], the model output is [62.34, 33.66], the error mae of the label recognition result is [0.16, 0.02], and the recognition error mae of all samples in the dataset D is shown in Table 1.

[0126] Table 1 Error table of recognition results

[0127]

[0128] As shown in Table 1, the mean absolute error (mae) of the rotational speed identification under the trained working conditions is 2.79 rpm, and the mae of the load identification is 6.82 N·m. Under the untrained working conditions, the mae of the rotational speed identification is 28.78 rpm, and the mae of the load identification is 54.59 N·m.

[0129] The present invention proposes a method for constructing a continuous working condition decoupling identification model based on a graph self-attention network. When a new sample is input into the saved working condition identification model, specific rotational speed and load values of the current signal will be given, realizing the decoupling and accurate identification of multiple working condition parameters.

Claims

1. A decoupling recognition method for continuous operating conditions of a diesel engine based on a graph self-attention network, characterized by including the following steps: The first step: Establish a sample data set 1.1 Collect the experimental data of diesel engine continuous operating condition changes according to the operating condition distribution, and sort out the total sample set D: D = {F 1 , F 2 , …, F t , …, F T} (1) in, D is the total sample set, F t represents the sample set of the t-th type of operating condition parameters, T is the number of operating condition categories, represents the i-th sample under the t-th type of operating condition parameters, n represents the number of samples under the t-th type of operating condition parameters, represents the j-th sequence point of the i-th sample under the t-th type of operating condition parameters, represents the K-th sequence point of the i-th sample under the t-th type of operating condition parameters, K is the number of sequence points included in each sample, and numerically equals the number of sequence points collected by the sensor during a complete working cycle of the machine operation; 1.2 Normalize the signal after filtering and noise reduction. Let x refer to a sample. That is, update each sample x as follows: f(x): x → x new (4) where x new is the normalized sample, x new ∈[-1,1], x min is the minimum value in the sample x sequence, x max is the maximum value in the sample x sequence; 1.3 Divide the total sample set D into a training set and a test set according to the sample category, with a ratio of 3:1, and the working condition category corresponding to the training set samples is different from that of the test set; Step 2: Design sample condition labels Due to the different dimensions of the working condition parameters, the accuracy of the output errors of different channels of the network will be different. Two independent parameters are used to represent the sample working condition labels: γ = [α, β], which are normalized by the rated speed and rated load respectively. The calculation process is as follows: Among them, V is the current rotational speed of the signal, and V idling is the idle speed of the diesel engine, and V rated is the rated rotational speed of the diesel engine; L is the current load of the signal, and L min is the minimum load of the diesel engine, and L min = 0, and L rated is the rated load of the diesel engine, and L rated The calculation process is as follows: where P rated is the rated power of the diesel engine; Step 3: Build a self-attention mechanism network model There is an intrinsic correlation between the latent features of signal fragments in different angular domains. For the same classification task, the latent features contain different classification contributions and have different effects on the recognition results. The self-attention mechanism is used to mine the key information in the signals of each time interval. At the input layer, the sample x is divided into m segments of angular domain signal segments with a dimension of d. The update process is as follows: x = [x 1 , x 2 ,..., x i ,..., x m T ∈R m×d (9)​ where x i is the i-th segment of the sample x; Calculate the Queries vector and Keys vector as follows: Among them, W q and are linear transformation matrices with dimensions of d×d k The dimensions of the Queries vector and the Keys vector are m×d k Here, the Queries vector represents the feature expression of the current signal segment, and the Keys vector represents the feature expressions of other signal segments in the input. Calculate the similarity between the Queries vector and all Keys vectors, and then apply an activation function to obtain the self-attention relationship matrix Att of the signal segment: where Softmax(·) is the softmax activation function, and d k is the dimension of the Queries vector and the Keys vector; Step 4: Build a graph convolutional network model Create graph data G. The creation process is as follows: G=(x,Att) (13) Among them, the angular domain signal fragment x is taken as a node, with a total of m nodes, and the self-attention coefficient Att is taken as an edge. The nodes in the graph aggregate the features of neighboring nodes through the graph convolution algorithm to generate new node representations. The update process of the graph convolution algorithm is as follows: h = σ(L sym xW h ) (14) where h is the output matrix of the graph convolutional layer, σ(·) is the non-linear activation function, L sym is the Symmetric normalized Laplacian matrix, W h is the trainable weight matrix with dimension d×d, is the adjacency matrix after adding self-loops, I is the identity matrix, Dig = diag(d(v 1 ), d(v 2 ), …, d(v n )) ∈ R n×n is the degree matrix of G, where d(v 1 ) = ∑ j∈V a ij is the degree of node v 1 , is 's degree matrix. The feature matrix x completes the process of propagating and receiving features between neighbor nodes by left-multiplying the adjacency matrix, and then right-multiplies a trainable weight matrix W h to complete the linear transformation, and finally is activated by the non-linear activation function to complete feature aggregation; Finally, the signal fragments are arranged and combined in the original order to obtain the signal enhancement matrix D′: X′=Reshape(h)∈R 1×K (16) D′ = {F′ 1 , F′ 2 ,..., F′ T} (18) Among them, Reshape() is a matrix dimension transformation function, X′ is the output vector after the dimension transformation of a sample, and F′ t is the dataset of the t-th type of working condition samples after signal enhancement, where is the first sample under the t-th type of working condition after signal enhancement, n represents the number of samples under the t-th type of working condition parameters, K represents the number of sequence points included in each sample, which is numerically equal to the number of sequence points collected by the sensor during a complete working cycle of the mechanical operation, D′ is the total sample dataset after signal enhancement, and T is the number of working condition categories; Step 5: Establish working condition decoupling space The network model classifier is defined as the working condition decoupling space, which consists of a fully connected layer with a node number equal to 2, namely the speed and load channels respectively, and the activation function is ReLu; the working condition decoupling space quantifies the working condition label, takes the speed and load channels as the reference axis, and uses the specific parameter value as the scale, and reduces the dimension of the sample data to the working condition decoupling space, visualizes the sample working condition identification value, and completes the decoupling and continuous identification of the speed and load parameters; Step 6: Establish the final working condition decoupling identification model; The final recognition model is established as a dual-channel model, each channel consists of a self-attention network layer, a graph convolution network layer and a working condition decoupling space classifier; the final model input is the sample after data preprocessing in the above steps 1.1, 1.2 and 1.3, and the output is the label of the sample in the above second step. The model goal is to minimize the mean absolute error mae of the predicted value, the loss function loss selects mae, and the optimization method selects Adam; finally, the working condition recognition model is trained, and the training is stopped when the mae gradient drops to 0, and the model is saved; Step 7: Decoupling and identification of operating condition parameters; When the machine is in the subsequent working process, the signal of a working cycle is arbitrarily extracted as a sample. After normalization in step 1.2, it is input into the model saved in step 6. The speed and load parameter values ​​of the sample are automatically displayed in the working condition decoupling space, and the value is a continuous value.

Images