A method for fault diagnosis of mechanical equipment based on vibration data

Abstract

The present invention provides a mechanical equipment fault diagnosis method based on vibration data, comprising the steps of: S1. Simultaneously collect multiple types of vibration data during mechanical equipment failures, and obtain a sampling data set by sampling according to time sequence; S2. Based on the sampling data set, generate a first training set and a first test set according to a set step size, set labels representing fault types for the first training set, and train a double-layer LSTM neural network for judging the fault types of mechanical equipment failures through the first training set; S3. Generate a second training set and a second test set based on the sampling data set, set labels representing fault types for the second training set, and train a BP neural network for judging the fault types of mechanical equipment through the second training set; S4. Compare the judgment accuracies of the two neural networks, select the one with a higher judgment accuracy as the actual fault judgment model, and judge the fault type based on the vibration data collected in real time through the actual fault judgment model.

Description

Technical Field

[0001] The present invention relates to the field of enterprise production automation management, and in particular to a mechanical equipment fault diagnosis method based on vibration data. Background Art

[0002] Time series working data is working data with time tags of mechanical equipment. Its typical characteristics are fast generation frequency, reliance on collection time, multiple measuring points and large amount of information. When enterprises carry out automation management, in order to ensure the safe, stable and efficient operation of equipment, they will also collect time series working data of mechanical equipment, such as temperature, humidity, pressure, vibration, and stress information of mechanical equipment through temperature, humidity, pressure, vibration, and strain sensors. At the same time, enterprises will also specially record the temperature and noise data of the factory. However, when mechanical equipment fails, the recorded time series working data is not fully utilized. This is mainly because the collected data is relatively complex, the data processing method is backward, and the processing of data requires additional resources, manpower and time. Many companies prefer to use traditional methods to diagnose equipment failures through maintenance personnel. However, in the traditional equipment fault diagnosis method, the location of the fault is basically based on the experience of the maintenance personnel for preliminary location, and then further testing is performed to accurately locate the fault problem. The traditional equipment fault diagnosis method has human subjective errors to a certain extent and prolongs the equipment maintenance time.

[0003] There are also methods to determine the type of equipment failure through simple modeling. The difficulty of this method lies in the multi-dimensional and non-steady-state data collection processing is more complicated, and it is necessary to set parameters based on expert experience, and it also relies on the experience and judgment of engineers to perform data analysis. On the other hand, noise interference is inevitable in the collected data. These noises will affect the correct diagnosis of mechanical failures, and it is difficult to explore the coupling relationship between the collected data and the health status of the equipment under low signal-to-noise ratio conditions.

[0004] Neural network models improve the accuracy and reliability of mechanical vibration analysis. However, directly importing undenoised data into a neural network model may lead to overfitting or underfitting of the model, affecting the accuracy of the model output results. In addition, shallow neural network models are prone to falling into local minima, gradient explosion and gradient vanishing problems, and poor model robustness.

[0005] How to quickly, efficiently and timely process the massive amount of time-series work data collected from production equipment has always been a major issue facing enterprise automation management.

[0006] Therefore, there is a need for an automated mechanical equipment fault diagnosis method that can evaluate the equipment operating status, perform reliability analysis on the equipment, and accurately locate equipment faults based on real-time collected production equipment working data. Summary of the Invention

[0007] The object of the present invention is to provide a method for diagnosing mechanical equipment faults based on vibration data, which can accurately and effectively identify the faults of mechanical equipment based on the vibration data of mechanical equipment, getting rid of the dependence on manual judgment of the types of mechanical equipment faults and greatly improving the work efficiency.

[0008] In order to achieve the above object, the present invention provides a method for diagnosing mechanical equipment faults based on vibration data, comprising the steps of:

[0009] S1. Simultaneously collect multiple types of vibration data when the mechanical equipment fails and denoise them, and sample the multiple types of vibration data according to time sequence to obtain a sampling data set;

[0010] S2. Based on the set time step, generate a first training set and a first test set from the sampling data set; manually set labels representing the types of mechanical equipment faults for the first training set; train a double-layer LSTM neural network through the first training set and the first test set; the double-layer LSTM neural network is used to judge the types of mechanical equipment faults; the first test set is used to verify the trained LSTM neural network;

[0011] S3. Generate a second training set and a second test set from the sampling data set; manually set labels representing the types of mechanical equipment faults for the second training set; train a BP neural network through the second training set, and the BP neural network is used to judge the types of mechanical equipment faults; the second test set is used to verify the trained BP neural network;

[0012] S4. Compare the judgment accuracies of the double-layer LSTM neural network and the BP neural network, and select the one with the higher judgment accuracy as the actual fault judgment model; judge the type of mechanical equipment fault based on the real-time collected vibration data of the mechanical equipment through the actual fault judgment model.

[0013] Optionally, the sampling data set is denoted as E, E = {e r} r∈[1,num] ; num is the total number of samplings; e r = {e′ r1 , …, e′ rm}; e′ rp is the p-th type of vibration data in the r-th sampling, p ∈ [1, m], and m is the total number of types of vibration data.

[0014] Optionally, the method for generating the first training set and the first test set in step S2 includes:

[0015] Let x i = [e (i-1)×s+1 , e(i-1)×s+2 , …, e i×s ′, where \(i\in[1, num / s]\), \(s\) is the set time step, and \([\cdot]'\) represents the transpose of a matrix; taking \(x\) 1 ~\(x\) L as the first training set, and taking as the first test set, \(L\) is a set constant.

[0016] Optionally, the double - layer LSTM network includes: a first hidden layer, a second hidden layer, and a flattening module;

[0017] When the double - layer LSTM network is trained for the \(i\) - th time, \(i\in[1, L - w + 1]\), it includes:

[0018] Inputting the time - series data \(x\) i , …, \(x\) i+w in the first test set into the first hidden layer in sequence. After \(w\) time - step operations by the LSTM standard module in the first hidden layer, the output results corresponding to the time steps are obtained where \(x\) i , …, \(x\) i+w are respectively used as the inputs of the LSTM standard module in the first hidden layer at these \(w\) time steps;

[0019] Inputting into the second hidden layer in sequence. After \(w\) time - step operations by the LSTM standard module in the second hidden layer, the output results corresponding to the time steps are obtained where are respectively used as the inputs of the LSTM standard module in the second hidden layer at these \(w\) time steps;

[0020] Inputting into the flattening module to obtain the classification result of the mechanical equipment fault type. The flattening module includes a classifier.

[0021] Optionally, the operation of the LSTM standard module for one time step includes:

[0022] S21. Denote the number of hidden units of the LSTM standard module as \(n\), and \(s\) is the set time step; at the current time step \(t\), the time step \(t - 1\) is the previous time step of the time step \(t\); represents the input at the time step \(t\), \(h\) t-1 represents the hidden state at the time step \(t - 1\), The forget gate \(f\) of the LSTM standard module at the time step \(t\) t and the input gate \(i\) t are respectively:

[0023] \(f\) t =\(\sigma(W\) f \cdot[ht-1 , x t + b f );

[0024] i t = σ(W i · [h t-1 , x t + b i );

[0025] where f t and · represents matrix operation, σ is the sigmod activation function, W f and are weight parameters, b f and are bias parameters;

[0026] S22. Calculate the new state candidate at time step t

[0027]

[0028] where tanh is the activation function, is the weight parameter, is the bias parameter;

[0029] S23. Calculate the update gate C at time step t t and the output gate o t :

[0030]

[0031] o t = σ(W o · [h t-1 , x t + b o );

[0032] where C t and C t-1 are the update gates at time step t - 1, are weight parameters, is the bias parameter;

[0033] S24. h t is the hidden state of the LSTM standard module at time step t and also the output result of the LSTM standard module at time step t:

[0034] h t = o t * tanh(Ct )

[0035] Among them,

[0036] Optionally, step S3 includes:

[0037] S31. Take as the second training set, as the second test set; n 1 ∈ [1, num]; Let a be the number of training times, and the initial value of a is 1;

[0038] S32. The BP neural network has a three-layer structure, including an input layer, a hidden layer, and an output layer; the number of nodes in the input layer is m; the number of nodes in the output layer is N, and one node in the output layer corresponds to a type of mechanical equipment failure; the hidden layer has B nodes;

[0039] During the a-th training, the input of the j-th node in the hidden layer where j ∈ [1, B], w ij is the connection weight from the i-th node in the input layer to the j-th node in the hidden layer, θ j is the threshold of the j-th node in the hidden layer; e′ ai is the vibration data of the i-th category in the a-th sampling, which corresponds to a node in the input layer;

[0040] The output of the j-th node in the hidden layer is b j = g(S j ), where g(·) is the Sigmoid function;

[0041] The input of the k-th node in the output layer where k ∈ [1, N], w′ lk is the connection weight from the l-th node in the hidden layer to the k-th node in the output layer, θ′ k is the threshold of the k-th node in the output layer;

[0042] The output of the k-th node in the output layer is y k = g(L k );

[0043] S33. Update the weights of w ij , w′ lk based on the Adam optimization algorithm, and update a to a + 1; repeat steps S32 to S33. When a > n 1 , complete the training of the BP neural network, and stop updating w ij , w′ lk , and enter S34;

[0044] S34. Verify the trained BP neural network through the first test set.

[0045] Optionally, the multi-class vibration data includes: vibration displacement, vibration velocity, vibration acceleration, and vibration amplitude of mechanical equipment.

[0046] Optionally, the denoising in step S1 is: eliminating the noise of the vibration data through wavelet transform, including the steps of;

[0047] S11. Select a wavelet function, perform wavelet transform on the vibration data to obtain a corresponding set of wavelet decomposition coefficients;

[0048] S12. Select a wavelet threshold, and delete the wavelet decomposition coefficients higher than and lower than the wavelet threshold;

[0049] S13. Perform wavelet reconstruction on the remaining wavelet decomposition coefficients through inverse wavelet transform to obtain the corresponding denoised vibration data.

[0050] Optionally, step S1 further includes preprocessing the collected vibration data before denoising; the preprocessing includes any one or more of signal amplification, signal attenuation, and DC bias.

[0051] Optionally, the wavelet function is the Daubechies wavelet function.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] 1) The present invention comprehensively utilizes multi-class vibration data of mechanical equipment, can effectively, accurately, and automatically judge and identify the fault types of mechanical equipment, gets rid of the dependence on manual judgment of mechanical equipment fault types, and greatly improves work efficiency;

[0054] 2) The present invention selects the one with a high judgment accuracy rate as the actual fault judgment model by actually comparing the judgment results of the BP neural network and the double-layer LSTM neural network. This actual fault judgment model can optimally match the types of the collected vibration data, and greatly improves the accuracy rate of mechanical equipment fault detection.

[0055] 3) The present invention effectively removes the noise in the collected vibration data, and improves the accuracy and precision of fault judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solution of the present invention, the drawings required to be used in the description will be briefly introduced below. Obviously, the drawings in the following description are an embodiment of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts:

[0057] Figure 1Flow chart of the mechanical fault diagnosis method based on mechanical equipment vibration signals according to the present invention;

[0058] Figure 2 Schematic diagram of the double - layer LSTM neural network structure according to the present invention;

[0059] Figure 3 Schematic diagram of the LSTM standard module according to the present invention;

[0060] Figure 4 Schematic diagram of the BP neural network structure according to the present invention. Detailed implementation manners

[0061] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0062] The present invention provides a mechanical equipment fault diagnosis method based on vibration data, as Figure 1 shown, including the steps:

[0063] S1. Simultaneously collect multiple types of vibration data during mechanical equipment failures, pre - process and denoise the vibration data, and sample the multiple types of vibration data according to time sequence to obtain a sampling data set;

[0064] In the embodiments of the present invention, the multiple types of vibration data include: vibration displacement, vibration amplitude, vibration velocity, and vibration acceleration of mechanical equipment collected by displacement sensors, velocity sensors, and acceleration sensors. The sampling data set is denoted as E, E = {e r} r∈[1,num] ; num is the total number of samplings; e r = {e′ r1 , …, e′ rm}; e′ rp is the p - th type of vibration data in the r - th sampling, p ∈ [1, m], and m is the total number of types of vibration data. In this embodiment, m = 4.

[0065] The pre - processing in step S1 includes any one or more of signal amplification, signal attenuation, and DC bias.

[0066] The denoising in step S1 is: eliminating the noise of the vibration data through wavelet transform, including the steps;

[0067] S11. Select a wavelet function, perform wavelet transform on the vibration data to obtain a set of corresponding wavelet decomposition coefficients. In the embodiment of the present invention, the wavelet function is the Daubechies wavelet function.

[0068] S12. Select a wavelet threshold, and delete the wavelet decomposition coefficients higher than and lower than the wavelet threshold.

[0069] S13. Perform wavelet reconstruction on the remaining wavelet decomposition coefficients through wavelet inverse transform to obtain the corresponding denoised vibration data.

[0070] S2. Based on the set time step, generate a first training set and a first test set from the sampling data set; manually set labels representing the fault types of mechanical equipment for the first training set; train a double-layer LSTM (Long Short-Term Memory) neural network through the first training set and the first test set; the double-layer LSTM neural network is used to judge the fault types of mechanical equipment faults; the first test set is used to verify the trained LSTM neural network.

[0071] In this embodiment, the method for generating the first training set and the first test set in step S2 includes:

[0072] Let x i =[e (i-1)×s+1 , e (i-1)×s+2 , …, e i×s ′, where i ∈ [1, num / s], s is the set time step, and [·]′ represents the transpose of the matrix; take x 1 ~x L as the first training set, and take as the first test set, and L is a set constant.

[0073] In this embodiment, as Figure 2 shown, the double-layer LSTM network includes: a first hidden layer, a second hidden layer, and a flattening module; extract the sequence features (hidden states) of the first training set through the LSTM standard modules of the first hidden layer and the second hidden layer, and input the sequence features into the flattening module to establish a mapping from the sequence features to the fault types. As Figure 2 shown, the input of the double-layer LSTM network adopts a sliding window form, the sliding window size is w, and at time step t, the input within the corresponding sliding window U t includes x t to x t+w .

[0074] As Figure 2 shown, when training the double-layer LSTM network for the i-th time, i ∈ [1, L - w + 1], it includes:

[0075] The timing data x in the first test set i , …, x i+w are sequentially input into the first hidden layer. After w time steps of operation by the LSTM standard module in the first hidden layer, the output results corresponding to the time steps are obtained where x i , …, x i+w serve as the inputs of the LSTM standard module in the first hidden layer at these w time steps respectively;

[0076] The are sequentially input into the second hidden layer. After w time steps of operation by the LSTM standard module in the second hidden layer, the output results corresponding to the time steps are obtained where serve as the inputs of the LSTM standard module in the second hidden layer at these w time steps respectively;

[0077] The is input into the flattening module to obtain the classification result of the mechanical equipment failure type. The flattening module includes a classifier.

[0078] In this embodiment, when the double - layer LSTM network is trained for the i - th time, the output of the flattening module where g(·) is the Softmax function, and g(·) can also be any classifier. Through the Softmax function, the output values of multi - classification can be converted into a probability distribution ranging from [0, 1] and summing to 1. Finally, the failure type of the mechanical equipment is obtained through classification by the flattening module.

[0079] In this embodiment, the flattening module includes C input nodes and C output nodes. One input node corresponds to one output node, and each output node corresponds to a type of failure. In this embodiment, C = 4. The input values of the input nodes are mapped to the output values of the corresponding output nodes through the Softmax function.

[0080]

[0081] where z p is the input value of the p - th input node of the flattening module, and Softmax(z p ) is the output value of the p - th output node of the flattening module. The failure type corresponding to the output node with the maximum output value is selected as the failure type of the mechanical equipment.

[0082] Figure 2 In 1 ~x T respectively represent the inputs of the first hidden layer at the first to the T - th time steps;

[0083] Denote the hidden states of the first hidden layer and the second hidden layer at the \(j\)-th time step; \(j\in[1,T]\).

[0084] Denote the update gates of the first hidden layer and the second hidden layer at the \(j\)-th time step; \(j\in[1,T]\).

[0085] Are all zero matrices, Are all zero matrices.

[0086] In this embodiment, as Figure 3 shown, the operation of the LSTM standard module for one time step includes:

[0087] S21. Denote the number of hidden units of the LSTM standard module as \(n\), and \(s\) as the set time step; at the current time step \(t\), the time step \(t - 1\) is the previous time step of the time step \(t\); Denote the input at the time step \(t\), \(h\) t-1 Denote the hidden state at the time step \(t - 1\), The forget gate \(f\) of the LSTM standard module at the time step \(t\) t and the input gate \(i\) t Are respectively:

[0088] \(f\) t =\(\sigma(W\) f \(\cdot[h\) t-1 ,x\) t +b\) f );

[0089] \(i\) t =\(\sigma(W\) i \(\cdot[h\) t-1 ,x\) t +b\) i );

[0090] Where, \(f\) t and \(\cdot\) represents matrix operation, \(\sigma\) is the sigmod activation function, \(W\) f and are weight parameters, \(b\) f and are bias parameters; the forget gate \(f\) t determines which information in the update gate \(C\) t-1 at the previous time step will be forgotten; the input gate \(i\) t determines which information in \(x\) t and \(h\) t-1 will be retained;

[0091] S22. Calculate the new state candidate quantity at the time step \(t\)

[0092]

[0093] Among them, tanh is the activation function, is the weight parameter, is the bias parameter;

[0094] S23. Calculate the update gate C at time step t t and the output gate o t :

[0095]

[0096] o t = σ(W o · [h t-1 , x t + b o );

[0097] Among them, C t and C t-1 are the update gates at time step t - 1, are weight parameters, is the bias parameter; i t and are multiplied to select which information will be added to C at time step t t , while f t is multiplied by C t-1 to determine which information in C t-1 will be retained and which information should be discarded. Integrate the output result h t-1 at time step t - 1 and the input x t at time step t through the sigmod activation function to obtain the output gate o t ;

[0098] S24. h t is the hidden state of the LSTM standard module at time step t and also serves as the output result of the LSTM standard module at time step t:

[0099] h t = o t * tanh(C t )

[0100] Among them,

[0101] S3. Generate a second training set and a second test set based on the sampling data set; manually set labels representing the fault types of mechanical equipment for the second training set; train a BP (Back Propagation) neural network through the second training set, and the BP neural network is used to judge the fault types of mechanical equipment failures; the second test set is used to verify the trained BP neural network;

[0102] In this embodiment, step S3 includes:

[0103] S31. Take as the second training set, as the second test set; n 1 ∈[1, num]; let a be the number of training times, and the initial value of a is 1;

[0104] S32. As shown in Figure 4 , the BP neural network has a three-layer structure, including an input layer, a hidden layer, and an output layer; the number of nodes in the input layer is m; the number of nodes in the output layer is N, and one node in the output layer corresponds to one type of mechanical equipment failure; the hidden layer has B nodes;

[0105] As shown in Figure 4 , in the embodiment of the present invention, m = 4, and the input layer has 4 nodes; in this embodiment, there are specifically four types of fault types, so N = 4, and the output layer includes Y 1 ~Y 4 a total of 4 nodes, and one node corresponds to one type of mechanical equipment failure; B = 10, and the hidden layer includes 10 nodes.

[0106] As shown in Figure 4 , during the a-th training, e′ a1 , …, e′ am respectively correspond to the m nodes in the input layer of the BP neural network, and the input of the j-th node in the hidden layer where j ∈ [1, B], w ij is the connection weight from the i-th node in the input layer to the j-th node in the hidden layer, θ j is the threshold of the j-th node in the hidden layer; e′ ai is the i-th type of vibration data sampled at the a-th time, which corresponds to a node in the input layer;

[0107] The output of the j-th node in the hidden layer is b j = g(S j ), where g(·) is the Sigmoid function;

[0108] The input of the k-th node in the output layer where k ∈ [1, N], w′ lkis the connection weight from the l-th node in the hidden layer to the k-th node in the output layer, and θ' k is the threshold of the k-th node in the output layer;

[0109] The output y of the k-th node in the output layer k = g(L k );

[0110] S33. Update the weights of w ij and w' lk , update a to a + 1; repeat steps S32 to S33. When a > n 1 , complete the training of the BP neural network, and stop updating w ij and w' lk , and enter S34;

[0111] S34. Verify the trained BP neural network through the first test set.

[0112] S4. Compare the judgment accuracies of the double-layer LSTM neural network and the BP neural network, and select the one with the higher judgment accuracy as the actual fault judgment model; based on the actually collected vibration data of the mechanical equipment, judge the type of mechanical equipment fault through the actual fault judgment model.

[0113] The present invention comprehensively utilizes various types of vibration data of mechanical equipment, can effectively, accurately and automatically judge and identify the types of mechanical equipment faults, gets rid of the dependence on manual judgment of mechanical equipment faults, and greatly improves the work efficiency;

[0114] The present invention selects the one with the higher judgment accuracy as the actual fault judgment model by actually comparing the judgment results of the BP neural network and the double-layer LSTM neural network. This actual fault judgment model can optimally match the types of the collected vibration data, and greatly improves the judgment accuracy of mechanical equipment faults. The present invention effectively removes the noise in the collected vibration data and improves the accuracy and precision of fault judgment.

[0115] The solution provided by the present invention reduces the algorithm fitting degree and improves the accuracy and reliability of mechanical vibration analysis. It provides a cheap and efficient means for vibration analysis and monitoring of the working state for some mechanical vibration-related equipment in the factory. The installation and transformation are simple and convenient, and the actual vibration data state of the mechanical equipment can be reliably monitored through digital signal processing and deep learning technology.

[0116] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for diagnosing mechanical equipment faults based on vibration data, characterized in that, it includes the steps: S1. Simultaneously collect multiple types of vibration data during mechanical equipment faults and denoise them. Sample the multiple types of vibration data according to time sequence to obtain a sampling data set; The sampled data set is denoted as E, E = {e r} r∈[1,num] ; num is the total number of samplings; e r = {e r ′ 1 , …, e r ′ m}; e r ′ p is the vibration data of the p-th class in the r-th sampling, p ∈ [1, m], and m is the total number of classes of vibration data; S2. According to the set time step, generate a first training set and a first test set based on the sampling data set; Manually set labels representing the types of mechanical equipment faults for the first training set; Train a double-layer LSTM neural network through the first training set and the first test set; The double-layer LSTM neural network is used to judge the types of mechanical equipment faults; The first test set is used to verify the trained LSTM neural network; S3. Generate a second training set and a second test set based on the sampling data set; Manually set labels representing the types of mechanical equipment faults for the second training set; Train a BP neural network through the second training set. The BP neural network is used to judge the types of mechanical equipment faults; The second test set is used to verify the trained BP neural network; Step S3 includes: S31. Take as the second training set, as the second test set; n 1 ∈ [1, num]; Let a be the number of training times, and the initial value of a is 1; S32. The BP neural network has a three-layer structure, including an input layer, a hidden layer, and an output layer; The number of nodes in the input layer is m; The number of nodes in the output layer is N, and one node in the output layer corresponds to one type of mechanical equipment fault; There are B nodes in the hidden layer; During the a-th training, the input of the j-th node in the hidden layer where j ∈ [1, B], w ij is the connection weight from the i-th node in the input layer to the j-th node in the hidden layer, and θ j is the threshold of the j-th node in the hidden layer; e′ ai is the vibration data of the i-th type in the a-th sampling, corresponding to a node in the input layer; The output of the j-th node in the hidden layer is b j = g(S j ), where g(·) is the Sigmoid function; Input to the k-th node in the output layer where k ∈ [1, N], w l ′ k is the connection weight from the l-th node in the hidden layer to the k-th node in the output layer, and θ k ′ is the threshold of the k-th node in the output layer; The output y of the k-th node in the output layer k = g(L k ); S33. Update w based on the Adam optimization algorithm ij and the weight of w l ′ k , update a to a + 1; repeat steps S32 to S33. When a > n 1 , complete the training of the BP neural network, and stop updating w ij and the weight of w l ′ k , enter S34; S34. Verify the trained BP neural network through the first test set; S4. Compare the judgment accuracies of the double-layer LSTM neural network and the BP neural network, and select the one with the higher judgment accuracy as the actual fault judgment model; Based on the real-time collected vibration data of the mechanical equipment, judge the type of mechanical equipment fault through the actual fault judgment model.

2. The method for diagnosing mechanical equipment faults based on vibration data according to claim 1, characterized in that, the method for generating the first training set and the first test set in step S2 includes: Let \(x\) i = [e (i-1)×s+1 , e (i-1)×s+2 , …, e i×s ', where \(i\in[1, num / s]\), \(s\) is the set time step, and \([\cdot]'\) represents the transpose of a matrix; Take \(x\) 1 ~ \(x\) L as the first training set, and take as the first test set, and \(L\) is the set constant.

3. The method for diagnosing mechanical equipment faults based on vibration data according to claim 2, characterized in that, the double-layer LSTM neural network includes: a first hidden layer, a second hidden layer, and a flattening module; When the double-layer LSTM neural network is trained for the i-th time, i ∈ [1, L - w + 1], it includes: Input the time series data x in the first test set i ,…,x i+w into the first hidden layer in sequence. After performing operations for w time steps through the LSTM standard module in the first hidden layer, obtain the output results corresponding to the time steps where x i ,…,x i+w serve as the inputs of the LSTM standard module in the first hidden layer at these w time steps respectively; Input sequentially into the second hidden layer, and after performing operations for w time steps through the LSTM standard module of the second hidden layer, obtain the output results corresponding to the time steps wherein are respectively used as the inputs of the LSTM standard module of the second hidden layer at these w time steps; respectively serve as the inputs of the LSTM standard module of the second hidden layer at these w time steps Input into the flattening module to obtain the classification result of the mechanical equipment fault type, where the flattening module includes a classifier.

4. The method for diagnosing mechanical equipment faults based on vibration data according to claim 3, characterized in that, the operation of the LSTM standard module for one time step includes: S21. Denote the number of hidden units in the standard LSTM module as n, and s as the set time step; at the current time step t, time step t - 1 is the previous time step of time step t; denotes the input at time step t, h t-1 denotes the hidden state at time step t - 1, the forget gate f of the standard LSTM module at time step t t and the input gate i t are respectively: f t = σ(W f · [h t-1 , x t + b f ); i t = σ(W i · [h t-1 , x t + b i ); Among them, f t and · represents matrix operation, σ is the sigmod activation function, W f and are weight parameters, b f and are bias parameters; S22, calculate the new state candidate quantity at time step t Among them, tanh is the activation function, is the weight parameter, is the bias parameter; S23. Calculate the update gate C at time step t t and the output gate o t : o t = σ(W o · [h t-1 , x t + b o ); Among them, C t and C t-1 is the update gate at time step t - 1, are weight parameters, is the bias parameter; S24, h t Is the hidden state of the LSTM standard module at time step t and also serves as the output result of the LSTM standard module at time step t: h t = o t *tanh(C t ) Among them, 5. The method for diagnosing mechanical equipment faults based on vibration data according to claim 1, characterized in that, the multiple types of vibration data include: vibration displacement, vibration velocity, vibration acceleration, and vibration amplitude of the mechanical equipment.

6. The method for diagnosing mechanical equipment faults based on vibration data according to claim 1, characterized in that, the denoising in step S1 is: through wavelet transform, eliminate the noise of the vibration data, including the steps; S11. Select a wavelet function, perform wavelet transform on the vibration data to obtain a corresponding set of wavelet decomposition coefficients; S12. Select a wavelet threshold, and delete the wavelet decomposition coefficients higher than and lower than the wavelet threshold; S13. Perform wavelet reconstruction on the retained wavelet decomposition coefficients through inverse wavelet transform to obtain the corresponding denoised vibration data.

7. The mechanical equipment fault diagnosis method based on vibration data according to claim 1, characterized in that, step S1 further includes preprocessing the collected vibration data before denoising; the preprocessing includes any one or more of signal amplification, signal attenuation, and DC bias.

8. The mechanical equipment fault diagnosis method based on vibration data according to claim 6, characterized in that, the wavelet function is the Daubechies wavelet function.

Images