System and method for multi-sensor information fusion fault diagnosis of rolling bearing

Abstract

The invention discloses a multi-sensor information fusion fault diagnosis system and method for a rolling bearing, and the method comprises the steps: continuously collecting sensor signals of M sensors at the same time, and carrying out the division according to a time period, thereby obtaining (I + J) samples, dividing the multi-sensor information fusion network into a training set # imgabs0 # and a verification set # imgabs1 # to construct a multi-sensor information fusion network structure search space SA = SO * SP and a multi-sensor information fusion network structure parameter alpha A = {alpha O, alpha P} for controlling the multi-sensor information fusion network structure search space SA, simultaneously updating a multi-sensor information fusion network structure parameter alpha A = {alpha O, alpha P} and a weight parameter w of the obtained neural network structure search space SA by adopting a training set # imgabs2 # and a verification set # imgabs3 #, and simultaneously searching a feature extraction unit and a fusion starting layer to obtain an optimal multi-sensor fusion network structure SA *, the calculation amount in the network search process is reduced, the efficiency is greatly improved, and the problem that the design of a current rolling bearing multi-sensor information fusion fault diagnosis method depends on artificial experience and the process is tedious due to a large number of experiment trials and errors is solved.

Description

Technical Field

[0001] The invention belongs to the field of rolling bearing fault diagnosis, and relates to a fault detection method, in particular to a system and method for multi-sensor information fusion fault diagnosis of rolling bearings. Background Art

[0002] Rolling bearings are the "joints" of mechanical equipment, playing a key role in bearing loads and transmitting motion. Once a bearing fails, it will affect the safe operation of the entire equipment and even cause immeasurable disasters. Therefore, it is of great significance to monitor and diagnose the bearing status. Traditional methods based on a single sensor, such as vibration sensors, are difficult to obtain more comprehensive bearing status information. As bearings serve in more complex and changeable working conditions, their diagnostic performance is severely limited. Multi-sensor information fusion technology can fully coordinate the complementary information from each sensor to enrich the available fault-related information, and is expected to achieve more accurate and reliable diagnostic performance in complex environments.

[0003] At present, the existing multi-sensor information fusion rolling bearing fault diagnosis is mostly based on deep neural networks, and the network structure, including feature extraction and fusion strategies, is mostly determined based on manual experience and a large number of experimental trial and error. For each different bearing model and different working condition diagnosis scenario, a targeted network structure must be redesigned. The network structure designed in this way is not necessarily the best. At the same time, the network structure of the multi-sensor fusion network is more complex than that of the single sensor network, with more parameters, which further expands its limitation of cumbersome development process and limits the engineering application of the multi-sensor information fusion rolling bearing fault diagnosis network. Summary of the invention

[0004] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a system and method for multi-sensor information fusion fault diagnosis of rolling bearings, so as to solve the problem that the current design of multi-sensor information fusion fault diagnosis method for rolling bearings relies on manual experience and a large number of cumbersome trial and error processes.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions to achieve the above problems:

[0006] A method for multi-sensor information fusion fault diagnosis of rolling bearings, specifically comprising the following steps:

[0007] Step 1: Simultaneously and continuously collect the sensor signals of M sensors and divide them into (I+J) samples according to the time period, which are divided into the training set and validation set

[0008] in:

[0009] (Xi trn ,Y i trn ) i represents the i-th sample in the training set, each of which includes the sensor signals and their labels of M sensors in a time period;

[0010] (X j val ,Y j val ) j represents the jth sample in the validation set, each of which includes the sensor signals and their labels of M sensors in a time period;

[0011] X i trn is the sensor signal of the M sensors in the i-th sample in the training set, m is the serial number of the sensor

[0012] X j val is the sensor signal of the M sensors in the jth sample in the validation set, m is the serial number of the sensor;

[0013] is the sensor signal of the mth sensor in the ith sample;

[0014] is the sensor signal of the mth sensor in the jth sample;

[0015] Y i trn is the label of the sensor signal of the M sensors in the i-th sample, and is the bearing fault category;

[0016] Y j val is the label of the sensor signal of the M sensors in the jth sample, which is the bearing fault category;

[0017] Step 2: Construct a multi-sensor information fusion network structure search space S A =S O ×S P and a search space S for controlling the multi-sensor information fusion network structure A The multi-sensor information fusion network structure parameter α A ={α O ,α P}, construct rolling bearing classifier h(·);

[0018] Among them, S O represents the feature extraction unit search space, S Prepresents the fusion strategy search space, α O Represents the search space S used to control the feature extraction unit O The structural parameter, α P Represents the search space S used to control the fusion strategy P Structural parameters of

[0019] Step 3: Use the training set and validation set At the same time, the neural network structure search space S obtained in step 2 A The multi-sensor information fusion network structure parameter α A ={α O ,α P} and weight parameter w are updated to obtain the optimal multi-sensor fusion network structure

[0020]

[0021]

[0022] in:

[0023] Represents the network structure parameter α from the multi-sensor information fusion network A ={α O ,α P} feature extraction unit search space S A Use search strategies in Extract a possible feature extraction unit network structure S;

[0024] w represents the weight parameter of the possible feature extraction unit network structure S;

[0025] To use the validation set D val The classification cross entropy loss of

[0026] To use the training set D trn The classification cross entropy loss of

[0027] Step 4: Collect multi-sensor signals in the scene to be diagnosed and input them into the optimal multi-sensor fusion network structure The rolling bearing fault diagnosis results are obtained

[0028] The present invention also includes the following technical features:

[0029] The steps 3 and 4 also include:

[0030] Using the training set Get the optimal multi-sensor fusion network structure for step 3 Conduct training.

[0031] In the above technical solution, since the purpose of step 3 is to obtain the optimal structure, the intermediate weight parameters are not necessarily optimal, so the training set is used. Get the optimal multi-sensor fusion network structure for step 3 Training is to train the weight parameters under the condition of fixing the overall structural parameters.

[0032] Step 2 specifically includes the following steps:

[0033] Step 2.1: Construct the feature extraction unit search space and control feature extraction unit search space S O The structural parameters are

[0034]

[0035]

[0036] in:

[0037] Cell m (N,O) represents the mth possible feature extraction unit based on the differentiable structure search unit;

[0038] o n,o' (·) represents the o'th operation on the nth edge;

[0039] N represents the edges that need to be retained in the differentiable structure search unit;

[0040] Represents the possible candidate operations set for each edge that is ultimately retained;

[0041] ||N|| indicates the number of edges that need to be retained in the end;

[0042] ||O|| indicates the number of edges that need to be retained in the end;

[0043] Represents the structural parameters of the o'th operation on the nth edge in the mth possible feature extraction unit;

[0044] thred indicates the operation retention threshold;

[0045] Step 2.2: Constructing the fusion strategy search space And control fusion strategy search space S P Structural parameters The fusion strategy search space Contains K layers of fusion modules connected in series

[0046] Step 2.3, search space for feature extraction units constructed according to step 2.1 and control feature extraction unit search space S O The structural parameters are The fusion strategy search space S constructed in step 2.2 P The control fusion strategy search space S P The structural parameters a P , and obtain the multi-sensor information fusion network structure search space S A =S O ×S P and multi-sensor information fusion network structure parameter α A ={α O ,α P}.

[0047] Step 3 specifically includes the following steps:

[0048] Step 3.1: Randomly initialize the multi-sensor information fusion network structure parameter α obtained in step 2 A ={α O ,α P};

[0049] Step 3.2, according to the multi-sensor information fusion network structure parameter α A ={α O ,α P}From the multi-sensor information fusion network structure search space S A Sample a possible feature extraction unit network structure S;

[0050] Step 3.3, initialize the weight parameters of the possible feature extraction unit network structure S obtained in step 3.2 w ;

[0051] Step 3.4, the training set obtained from step 1 and validation set The i-th sample (X trn ,Y trn ) i and the jth sample (X val ,Y val ) j ;

[0052] Step 3.5, use the i-th sample (X trn ,Y trn ) i Update the weight parameter w with the learning rate γ1 to obtain the updated weight parameter w;

[0053] Step 3.6, let i=i+1, repeat steps 3.4 to 3.5 until Equal to zero or i = I, get the optimal network weight parameter w * ;

[0054] Step 3.7, use the jth sample (X val ,Y val ) j The learning rate γ2 is used to control the search space S of the feature extraction unit O The structural parameter α O Update to obtain the updated search space S for controlling the feature extraction unit O The structural parameter α O ;

[0055] Step 3.8, let j = j + 1, repeat steps 3.2 to 3.7 until Equal to zero or j = J, get the updated neural network structure search space S A The multi-sensor information fusion network structure parameter α A ={α O ,α P}, that is, the optimal multi-sensor fusion network structure is obtained

[0056] Step 3.5 specifically includes the following steps:

[0057] Step 3.5.1: Take the i-th sample (X trn ,Y trn ) i middle The mth sensor information Input the corresponding encoder Perform feature transformation to obtain encoded features

[0058] Step 3.5.2: Encoded features Input to the fusion strategy search space Fusion and output fusion features

[0059] Step 3.5.3: The fusion features obtained in step 3.5.2 Input to rolling bearing classifier h(·), output diagnosis result And calculate the diagnosis result Y i 'and The cross entropy loss

[0060] Step 3.5.4, using cross entropy loss The network weight parameter w is updated to obtain the updated weight parameter w.

[0061] Step 3.5.2 specifically includes the following steps:

[0062] Step 3.5.2.1, initialize k=1;

[0063] Step 3.5.2.2, determine if k≤2, if yes, proceed to step 3.5.2.3, otherwise proceed to step 3.5.2.6;

[0064] Step 3.5.2.3, determine if k=1, if yes, proceed to step 3.5.2.4, otherwise proceed to step 3.5.2.5;

[0065] Step 3.5.2.4, the encoded features are obtained The elements in are concatenated and copied into two identical copies as the k-th layer fusion module The first input and the second input Right now Follow step 3.5.2.7;

[0066] in:

[0067] Represents a splicing operation;

[0068] represents the output of the k-2th layer fusion module;

[0069] Step 3.5.2.5, the encoded features are obtained The elements in are concatenated as a k-layer fusion module The second input Follow step 3.5.2.7;

[0070] Step 3.5.2.6, merge the k-2th layer of modules and the k-1th layer fusion module Output and As the k-th layer fusion module For the first and second inputs, execute step 3.5.2.7;

[0071] Step 3.5.2.7, search space through feature extraction unit For the k-th layer fusion module The first input and the second input Perform feature extraction to obtain feature sets

[0072] in:

[0073] Indicates that the i-th sample is fused in the k-th layer module In the example, the sensor signal of the mth sensor passes through the corresponding mth feature extraction unit Cell m Output of (N,O);

[0074] and Respectively represent the information of the mth sensor of the i-th sample in the output of the k-2-layer fusion module and the k-1-layer fusion module;

[0075] Step 3.5.2.8, the feature set obtained in step 3.5.2.7 Concatenate the elements in to get

[0076] Where R represents the real number domain, r and d both represent the feature dimension, and r = M × l, l is the length of the feature map corresponding to each sensor signal;

[0077] Step 3.5.2.9, calculate the attention score matrix A without softmax operation by the following formula o ;

[0078]

[0079]

[0080] in:

[0081] Represents the intermediate process of calculating the self-attention score;

[0082] Q and K are the query matrix and key matrix;

[0083] Step 3.5.2.10, for the attention score matrix A o Superposition mask matrix M o ∈R r×d Get the masked attention score A m =A o +M o ;

[0084] The mask matrix M o There are two states: if k ≥ k f , that is, fusion is performed, then M o is a matrix of all 0s, that is, M o =0; if k <k f , that is, no fusion is performed, then M oThe corresponding position elements where different information sources interact are -∞, and the rest are 0, that is,

[0085]

[0086] in:

[0087] k f Indicates the fusion starting layer,

[0088] Step 3.5.2.11, use the masked attention score A m Perform calculations on the self-attention mechanism and get the output

[0089] Among them, softmax(·) represents the softmax function;

[0090] V is the value matrix;

[0091] Step 3.5.2.12, let k = k + 1, return to step 3.5.2.2, until k = K, output fusion features

[0092] A system for multi-sensor information fusion fault diagnosis of rolling bearings, based on the multi-sensor information fusion fault diagnosis method for rolling bearings, comprising a data acquisition unit, a model building unit and a fault detection unit;

[0093] The data acquisition unit is used to simultaneously and continuously collect sensor signals of M sensors, and obtain (I+J) samples according to time periods, which are divided into training sets and validation set

[0094] in:

[0095] (X i trn ,Y i trn ) i represents the i-th sample in the training set, each of which includes the sensor signals and their labels of M sensors in a time period;

[0096] (X j val ,Y j val ) j represents the jth sample in the validation set, each of which includes the sensor signals and their labels of M sensors in a time period;

[0097] X i trnis the sensor signal of the M sensors in the i-th sample in the training set, m is the serial number of the sensor

[0098] X j val is the sensor signal of the M sensors in the jth sample in the validation set, m is the serial number of the sensor;

[0099] is the sensor signal of the mth sensor in the ith sample;

[0100] is the sensor signal of the mth sensor in the jth sample;

[0101] Y i trn is the label of the sensor signal of the M sensors in the i-th sample, and is the bearing fault category;

[0102] Y j val is the label of the sensor signal of the M sensors in the jth sample, which is the bearing fault category;

[0103] The model building unit is used to build a multi-sensor information fusion network structure search space S A =S O ×S P and a search space S for controlling the multi-sensor information fusion network structure A The multi-sensor information fusion network structure parameter α A ={α O ,α P}, construct rolling bearing classifier h(·); where S O represents the feature extraction unit search space, S P represents the fusion strategy search space, α O Represents the search space S used to control the feature extraction unit O The structural parameter, α P Represents the search space S used to control the fusion strategy P Structural parameters of the training set and validation set At the same time, the neural network structure search space S obtained in step 2 A The multi-sensor information fusion network structure parameter α A ={α O ,α P} and weight parameter w are updated to obtain the optimal multi-sensor fusion network structure

[0104]

[0105]

[0106] in:

[0107] Represents the network structure parameter α from the multi-sensor information fusion network A ={α O ,α P} feature extraction unit search space S A Use search strategies in Extract a possible feature extraction unit network structure S;

[0108] w represents the weight parameter of the possible feature extraction unit network structure S;

[0109] To use the validation set D val The classification cross entropy loss of

[0110] To use the training set D trn The classification cross entropy loss of

[0111] The fault detection unit is used to collect multi-sensor signals in the scene to be diagnosed and input them into the optimal multi-sensor fusion network structure In the experiment, the rolling bearing fault diagnosis results are obtained.

[0112] Compared with the prior art, the present invention has the following beneficial technical effects:

[0113] (I) The present invention obtains the optimal multi-sensor fusion network structure by searching the feature extraction unit and the fusion start layer at the same time This reduces the amount of calculation in the network search process and greatly improves efficiency, solving the cumbersome problem that the current design of rolling bearing multi-sensor information fusion fault diagnosis methods relies on manual experience and a large number of experimental trial and error.

[0114] (II) Construct a fusion strategy search space composed of multiple fusion modules, extract features in the fusion module through the feature extraction unit, and use the masked self-attention mechanism to control the fusion, so that the entire network structure is more modular. Through the search strategy, a rolling bearing multi-sensor information fusion fault diagnosis network for different scenarios can be quickly developed.

[0115] (III) The optimal feature extraction unit and fusion strategy are adaptively determined through a data-driven search method, which overcomes the tedious and time-consuming development process of the existing rolling bearing multi-sensor information fusion fault diagnosis network, and the defects that the designed network is not necessarily optimal. Through the basic modules, a rolling bearing multi-sensor information fusion fault diagnosis network with optimal accuracy in different scenarios is quickly built, which effectively promotes the balance between diagnostic accuracy and development efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0116] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0117] Figure 2 The fusion process of the fusion strategy search space in the present invention;

[0118] Figure 3 It is a schematic diagram of the structure of the Differentiable Artificial Intelligence Search Unit (DARTS);

[0119] Figure 4 These are the feature extraction units searched out in this embodiment, including: (a) BPS data set signal 1 feature extraction unit; (b) BPS data set signal 2 feature extraction unit; (c) PU data set signal 1 feature extraction unit; (d) PU data set signal 2 feature extraction unit.

[0120] The specific contents of the present invention are further explained in detail below in conjunction with embodiments. DETAILED DESCRIPTION

[0121] It should be noted that, unless otherwise specified, all components in the present invention are components known in the art.

[0122] Specific embodiments of the present invention are given below. It should be noted that the present invention is not limited to the following specific embodiments, and all equivalent changes made on the basis of the technical solution of this application fall within the protection scope of the present invention.

[0123] The present invention provides a method for multi-sensor information fusion fault diagnosis of rolling bearings, which specifically comprises the following steps:

[0124] A method for multi-sensor information fusion fault diagnosis of rolling bearings, specifically comprising the following steps:

[0125] Step 1: Simultaneously and continuously collect the sensor signals of M sensors and divide them into (I+J) samples according to the time period, which are divided into the training set and validation set

[0126] in:

[0127] (Xi trn ,Y i trn ) i represents the i-th sample in the training set, each of which includes the sensor signals and their labels of M sensors in a time period;

[0128] (X j val ,Y j val ) j represents the jth sample in the validation set, each of which includes the sensor signals and their labels of M sensors in a time period;

[0129] X i trn is the sensor signal of the M sensors in the i-th sample in the training set, m is the serial number of the sensor

[0130] X j val is the sensor signal of the M sensors in the jth sample in the validation set, m is the serial number of the sensor;

[0131] is the sensor signal of the mth sensor in the ith sample;

[0132] is the sensor signal of the mth sensor in the jth sample;

[0133] Y i trn is the label of the sensor signal of the M sensors in the i-th sample, and is the bearing fault category;

[0134] Y j val is the label of the sensor signal of the M sensors in the jth sample, which is the bearing fault category;

[0135] Step 2: Construct a multi-sensor information fusion network structure search space S A =S O ×S P And the search space S for controlling the multi-sensor information fusion network structure A The multi-sensor information fusion network structure parameter α A ={α O ,α P}, construct rolling bearing classifier h(·);

[0136] Among them, S O represents the feature extraction unit search space, S Prepresents the fusion strategy search space, α O Represents the search space S used to control the feature extraction unit O The structural parameter, α P Represents the search space S used to control the fusion strategy P Structural parameters of

[0137] Step 3: Use the training set and validation set At the same time, the neural network structure search space S obtained in step 2 A The multi-sensor information fusion network structure parameter α A ={α O ,α P} and weight parameter w are updated to obtain the optimal multi-sensor fusion network structure S A* ;

[0138]

[0139]

[0140] in:

[0141] Represents the network structure parameter α from the multi-sensor information fusion network A ={α O ,α P} feature extraction unit search space S A Use search strategies in Extract a possible feature extraction unit network structure S;

[0142] w represents the weight parameter of the possible feature extraction unit network structure S;

[0143] To use the validation set D val The classification cross entropy loss of

[0144] To use the training set D trn The classification cross entropy loss of

[0145] Step 4: Collect multi-sensor signals in the scene to be diagnosed and input them into the optimal multi-sensor fusion network structure The rolling bearing fault diagnosis results are obtained

[0146] In the above technical solution, the optimal multi-sensor fusion network structure is obtained by searching the feature extraction unit and the fusion start layer at the same time. This reduces the amount of calculation in the network search process and greatly improves efficiency, solving the cumbersome problem that the current design of rolling bearing multi-sensor information fusion fault diagnosis methods relies on manual experience and a large number of experimental trial and error.

[0147] The sensor is a vibration sensor, a sound sensor, a rotation speed sensor, a torque sensor or a force sensor.

[0148] Between step 3 and step 4 also include:

[0149] Using the training set Get the optimal multi-sensor fusion network structure for step 3 Conduct training.

[0150] Step 2 specifically includes the following steps:

[0151] Step 2.1: Construct the feature extraction unit search space and control feature extraction unit search space S O The structural parameters are

[0152]

[0153] in:

[0154] Cell m (N,O) represents the mth possible feature extraction unit based on the differentiable structure search unit;

[0155] o n,o' (·) represents the o'th operation on the nth edge;

[0156] N represents the edges that need to be retained in the differentiable structure search unit;

[0157] Represents the possible candidate operations set for each edge that is ultimately retained;

[0158] ||N|| indicates the number of edges that need to be retained in the end;

[0159] ||O|| indicates the number of edges that need to be retained in the end;

[0160] Represents the structural parameters of the o'th operation on the nth edge in the mth possible feature extraction unit;

[0161] thred indicates the operation retention threshold;

[0162] Step 2.2: Constructing the fusion strategy search space And control fusion strategy search space S P Structural parameters Fusion strategy search space Contains K layers of fusion modules connected in series

[0163] Step 2.3, search space for feature extraction units constructed according to step 2.1 and control feature extraction unit search space S O The structural parameters are The fusion strategy search space S constructed in step 2.2 P The control fusion strategy search space S P The structural parameters a P , and obtain the multi-sensor information fusion network structure search space S A =S O ×S P and multi-sensor information fusion network structure parameter α A ={α O ,α P}.

[0164] In the above technical scheme, a fusion strategy search space composed of multiple fusion modules is constructed, features are extracted in the fusion module through the feature extraction unit, and the masked self-attention mechanism is used to control the fusion, so that the entire network structure is more modular. Through the search strategy, a rolling bearing multi-sensor information fusion fault diagnosis network for different scenarios can be quickly developed.

[0165] Step 3 specifically includes the following steps:

[0166] Step 3.1: Randomly initialize the multi-sensor information fusion network structure parameter α obtained in step 2 A ={α O ,α P};

[0167] Step 3.2, according to the multi-sensor information fusion network structure parameter α A ={α O ,α P}From the multi-sensor information fusion network structure search space S A Sample a possible feature extraction unit network structure S;

[0168] Step 3.3, initializing the weight parameter w of the possible feature extraction unit network structure S obtained in step 3.2;

[0169] Step 3.4, the training set obtained from step 1 and validation set The i-th sample (X trn ,Y trn ) i and the jth sample (Xval ,Y val ) j ;

[0170] Step 3.5, use the i-th sample (X trn ,Y trn ) i Update the weight parameter w with the learning rate γ1 to obtain the updated weight parameter w ;

[0171] Step 3.6, let i=i+1, repeat steps 3.4 to 3.5 until Equal to zero or i = I, get the optimal network weight parameter w * ;

[0172] Step 3.7, use the jth sample (X val ,Y val ) j The learning rate γ2 is used to control the search space S of the feature extraction unit O The structural parameter α O Update to obtain the updated search space S for controlling the feature extraction unit O The structural parameter α O ;

[0173] Step 3.8, let j = j + 1, repeat steps 3.2 to 3.7 until Equal to zero or j = J, get the updated neural network structure search space S A The multi-sensor information fusion network structure parameter α A ={α O ,α P}, that is, the optimal multi-sensor fusion network structure S is obtained A* .

[0174] In the above technical scheme, the optimal feature extraction unit and fusion strategy are adaptively determined through a data-driven search method, which overcomes the cumbersome and time-consuming development process of the existing rolling bearing multi-sensor information fusion fault diagnosis network, and the defects that the designed network is not necessarily optimal. The rolling bearing multi-sensor information fusion fault diagnosis network with optimal accuracy in different scenarios is quickly built through basic modules, which effectively promotes the balance between diagnostic accuracy and development efficiency.

[0175] In step 3.2, when the total number of operations ||O||=8, the operation retention threshold thred=0.5, the structural parameters of the feature extraction unit corresponding to the mth sensor are When The operations used by the feature extraction unit of α are the fifth and seventh operations; PUsed to control which layer the network starts to merge. For example, when K=5, the fusion starts at layer At this time, from the kth f The layers begin to merge. Thus, the network structure S can be uniquely determined.

[0176] Step 3.5 specifically includes the following steps:

[0177] Step 3.5.1: Take the i-th sample (X trn ,Y trn ) i middle The mth sensor information Input the corresponding encoder Perform feature transformation to obtain encoded features

[0178] Step 3.5.2: Encoded features Input to the fusion strategy search space Fusion and output fusion features

[0179] Step 3.5.3: The fusion features obtained in step 3.5.2 Input to rolling bearing classifier h(·), output diagnosis result And calculate the diagnosis result Y i 'With Y i trn The cross entropy loss

[0180] Step 3.5.4, using cross entropy loss The network weight parameter w is updated to obtain the updated weight parameter w.

[0181] Step 3.5.2 specifically includes the following steps:

[0182] Step 3.5.2.1, initialize k=1;

[0183] Step 3.5.2.2, determine if k≤2, if yes, proceed to step 3.5.2.3, otherwise proceed to step 3.5.2.6;

[0184] Step 3.5.2.3, determine if k=1, if yes, proceed to step 3.5.2.4, otherwise proceed to step 3.5.2.5;

[0185] Step 3.5.2.4, the encoded features are obtained The elements in are concatenated and copied into two identical copies as the k-th layer fusion module The first input and the second input Right now Follow step 3.5.2.7;

[0186] in:

[0187] Represents a splicing operation;

[0188] represents the output of the k-2th layer fusion module;

[0189] Step 3.5.2.5, the encoded features are obtained The elements in are concatenated as a k-layer fusion module The second input Follow step 3.5.2.7;

[0190] Step 3.5.2.6, merge the k-2th layer of modules and the k-1th layer fusion module Output and As the k-th layer fusion module For the first and second inputs, execute step 3.5.2.7;

[0191] Step 3.5.2.7, search space through feature extraction unit For the k-th layer fusion module The first input and the second input Perform feature extraction to obtain feature sets

[0192] in:

[0193] Indicates that the i-th sample is fused in the k-th layer module In the example, the sensor signal of the mth sensor passes through the corresponding mth feature extraction unit Cell m Output of (N,O);

[0194] and Respectively represent the information of the mth sensor of the i-th sample in the output of the k-2-layer fusion module and the k-1-layer fusion module;

[0195] Step 3.5.2.8, the feature set obtained in step 3.5.2.7 Concatenate the elements in to get

[0196] Where R represents the real number domain, r and d both represent the feature dimension, and r = M × l, l is the length of the feature map corresponding to each sensor signal;

[0197] Step 3.5.2.9, calculate the attention score matrix A without softmax operation by the following formula o ;

[0198]

[0199]

[0200] in:

[0201] Represents the intermediate process of calculating the self-attention score;

[0202] Q and K are the query matrix and key matrix;

[0203] Step 3.5.2.10, for the attention score matrix A o Superposition mask matrix M o ∈R r×d Get the masked attention score A m =A o +M o ;

[0204] Mask matrix M o There are two states: if k ≥ k f , that is, fusion is performed, then M o is a matrix of all 0s, that is, M o =0; if k <k f , that is, no fusion is performed, then M o The corresponding position elements where different information sources interact are -∞, and the rest are 0, that is,

[0205]

[0206] in:

[0207] k f Indicates the fusion starting layer,

[0208] Step 3.5.2.11, use the masked attention score A m Perform calculations on the self-attention mechanism and get the output

[0209] Among them, softmax(·) represents the softmax function;

[0210] V is the value matrix;

[0211] Step 3.5.2.12, let k = k + 1, return to step 3.5.2.2, until k = K, output fusion features

[0212] In the above technical solution, when M o When all elements in are 0, A m Perform normal attention score calculation and information fusion, while A m The elements with -∞ in ∈R become 0 after softmax(·), that is, the fused parts of different information sources are masked and the fusion fails.

[0213] The present invention also provides a system for multi-sensor information fusion fault diagnosis of rolling bearings, based on the method for multi-sensor information fusion fault diagnosis of rolling bearings, including a data acquisition unit, a model building unit and a fault detection unit;

[0214] The data acquisition unit is used to continuously collect the sensor signals of M sensors simultaneously and obtain (I+J) samples according to the time period, which are divided into training sets and validation set

[0215] in:

[0216] (X i trn ,Y i trn ) i represents the i-th sample in the training set, each of which includes the sensor signals and their labels of M sensors in a time period;

[0217] (X j val ,Y j val ) j represents the jth sample in the validation set, each of which includes the sensor signals and their labels of M sensors in a time period;

[0218] X i trn is the sensor signal of the M sensors in the i-th sample in the training set, m is the serial number of the sensor

[0219] X j val is the sensor signal of the M sensors in the jth sample in the validation set, m is the serial number of the sensor;

[0220] is the sensor signal of the mth sensor in the ith sample;

[0221] is the sensor signal of the mth sensor in the jth sample;

[0222] Y i trn is the label of the sensor signal of the M sensors in the i-th sample, and is the bearing fault category;

[0223] Y j val is the label of the sensor signal of the M sensors in the jth sample, which is the bearing fault category;

[0224] The model building unit is used to construct the multi-sensor information fusion network structure search space S A =S O ×S P And the search space S for controlling the multi-sensor information fusion network structure A The multi-sensor information fusion network structure parameter α A ={α O ,α P}, construct rolling bearing classifier h(·); where S O represents the feature extraction unit search space, S P represents the fusion strategy search space, α O Represents the search space S used to control the feature extraction unit O The structural parameter, α P Represents the search space S used to control the fusion strategy P Structural parameters of the training set and validation set At the same time, the neural network structure search space S obtained in step 2 A The multi-sensor information fusion network structure parameter α A ={α O ,α P} and weight parameter w are updated to obtain the optimal multi-sensor fusion network structure

[0225]

[0226]

[0227] in:

[0228] Represents the network structure parameter α from the multi-sensor information fusion network A ={α O ,α P} feature extraction unit search space S A Use search strategies in Extract a possible feature extraction unit network structure S;

[0229] w represents the weight parameter of the possible feature extraction unit network structure S;

[0230] To use the validation set D val The classification cross entropy loss of

[0231] To use the training set D trn The classification cross entropy loss of

[0232] The fault detection unit is used to collect multi-sensor signals in the scene to be diagnosed and input them into the optimal multi-sensor fusion network structure S A* In the experiment, the rolling bearing fault diagnosis results are obtained.

[0233] Example:

[0234] This embodiment provides a method for multi-sensor information fusion fault diagnosis of rolling bearings. This example uses the rolling bearing data set of Paderborn University and the BPS data set of the Key Laboratory of Modern Design and Rotor Bearing Systems of the Ministry of Education to verify the method of the present invention, as shown in Table 1:

[0235] The Paderborn University rolling bearing dataset contains the health status of bearing model 6203: normal, inner ring fault, and outer ring fault. The signals used in the experiment are force signals and torque signals, the working conditions are 1500rpm speed, 0.7N·m load, and 1kN radial force, and the sampling rate is 4kHz. The dataset is expanded to obtain 2000, 1000, and 1000 training datasets, 1000, and 1000 for each fault type, respectively.

[0236] The BPS data set contains the health status of the bearing model 7014C: normal, inner ring fault, outer ring fault, and rolling element fault. The signals used in the experiment are two orthogonal vibration signals 1 and 2, the working conditions are 900rpm speed, 1kN axial load, and the sampling rate is 32kHz. The data set is overlapped and the number of training data sets, validation data sets, and test data sets for each fault type are 1800, 200, and 200 respectively.

[0237] Table 1 Paderborn University rolling bearing dataset used in the experiment

[0238]

[0239]

[0240] Based on the data sets shown in Table 1, the diagnostic tasks of the two data sets are constructed. First, the searched feature extraction units are as follows: Figure 4 As shown, it can be seen that different structures are searched out in different scenarios. Then the fusion start layers searched out by the two tasks are the third layer and the fifth layer respectively. The accuracy comparison with other methods is shown in Table 2, where EDCNN is a method that uses two-way convolution to extract two signal features respectively and then fuses them. It has fixed structural parameters in the two tasks. MM-NAS and MFAS are two multimodal fusion search methods. It can be seen that EDCNN5 with a fixed structure lacks applicability in different tasks and has the worst performance. The proposed method achieves the highest diagnostic effect, indicating that it can automatically search for the optimal network structure suitable for different diagnostic tasks without a tedious design process.

[0241] Table 2 Comparison of accuracy of different methods

[0242]

Claims

1. A method for multi-sensor information fusion fault diagnosis of rolling bearings, characterized in that: The specific steps include: Step 1: Simultaneously and continuously collect sensor signals from M sensors and divide them into (I+J) samples according to time periods, which are divided into training sets and validation set in: represents the i-th sample in the training set, each of which includes the sensor signals and their labels of M sensors in a time period; represents the jth sample in the validation set, each of which includes the sensor signals and their labels of M sensors in a time period; is the sensor signal of the M sensors in the i-th sample in the training set, m is the serial number of the sensor is the sensor signal of the M sensors in the jth sample in the validation set, m is the serial number of the sensor; is the sensor signal of the mth sensor in the ith sample; is the sensor signal of the mth sensor in the jth sample; is the label of the sensor signal of the M sensors in the i-th sample, and is the bearing fault category; is the label of the sensor signal of the M sensors in the jth sample, which is the bearing fault category; Step 2: Construct a multi-sensor information fusion network structure search space S A =S O ×S P and a search space S for controlling the multi-sensor information fusion network structure A The multi-sensor information fusion network structure parameter α A ={α O ,α P }, construct rolling bearing classifier h(·); Among them, S O represents the feature extraction unit search space, S P represents the fusion strategy search space, α O Represents the search space S used to control the feature extraction unit O The structural parameter, α P Represents the search space S used to control the fusion strategy P Structural parameters of Step 3: Use the training set and validation set At the same time, the neural network structure search space S obtained in step 2 A The multi-sensor information fusion network structure parameter α A ={α O ,α P } and weight parameter w are updated to obtain the optimal multi-sensor fusion network structure S A* ; in: Represents the network structure parameter α from the multi-sensor information fusion network A ={α O ,α P } feature extraction unit search space S A Use search strategies in Extract a possible feature extraction unit network structure S; w represents the weight parameter of the possible feature extraction unit network structure S; To use the validation set D val The classification cross entropy loss of To use the training set D trn The classification cross entropy loss of Step 4: Collect multi-sensor signals in the scene to be diagnosed and input them into the optimal multi-sensor fusion network structure S A* The rolling bearing fault diagnosis results are obtained 2. The method for multi-sensor information fusion fault diagnosis of rolling bearings as claimed in claim 1, characterized in that: The steps 3 and 4 also include: Using the training set For step 3, we get the optimal multi-sensor fusion network structure S A* Conduct training.

3. The method for multi-sensor information fusion fault diagnosis of rolling bearings as claimed in claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2.1: Construct the feature extraction unit search space and control feature extraction unit search space S O The structural parameters are in: Cell m (N,O) represents the mth possible feature extraction unit based on the differentiable structure search unit; o n,o' (·) represents the o'th operation on the nth edge; N represents the edges that need to be retained in the differentiable structure search unit; Represents the possible candidate operations set for each edge that is ultimately retained; ||N|| indicates the number of edges that need to be retained in the end; ||O|| indicates the number of edges that need to be retained in the end; Represents the structural parameters of the o'th operation on the nth edge in the mth possible feature extraction unit; thred indicates the operation retention threshold; Step 2.2: Constructing the fusion strategy search space And control fusion strategy search space S P Structural parameters The fusion strategy search space Contains K layers of fusion modules connected in series Step 2.3, search space for feature extraction units constructed according to step 2.1 and control feature extraction unit search space S O The structural parameters are The fusion strategy search space S constructed in step 2.2 P The control fusion strategy search space S P The structural parameters a P , and obtain the multi-sensor information fusion network structure search space S A =S O ×S P and multi-sensor information fusion network structure parameter α A ={α O ,α P }.

4. The method for multi-sensor information fusion fault diagnosis of rolling bearings as claimed in claim 1, characterized in that: Step 3 specifically includes the following steps: Step 3.1: Randomly initialize the multi-sensor information fusion network structure parameter α obtained in step 2 A ={α O ,α P }; Step 3.2, according to the multi-sensor information fusion network structure parameter α A ={α O ,α P }From the multi-sensor information fusion network structure search space S A Sample a possible feature extraction unit network structure S; Step 3.3, initializing the weight parameter w of the possible feature extraction unit network structure S obtained in step 3.2; Step 3.4, the training set obtained from step 1 and validation set The i-th sample (X trn ,Y trn ) i and the jth sample (X val ,Y val ) j ; Step 3.5, use the i-th sample (X trn ,Y trn ) i Update the weight parameter w with the learning rate γ1 to obtain the updated weight parameter w; Step 3.6, let i=i+1, repeat steps 3.4 to 3.5 until Equal to zero or i = I, get the optimal network weight parameter w * ; Step 3.7, use the jth sample (X val ,Y val ) j The learning rate γ2 is used to control the search space S of the feature extraction unit O The structural parameter α O Update to obtain the updated search space S for controlling the feature extraction unit O The structural parameter α O ; Step 3.8, let j = j + 1, repeat steps 3.2 to 3.7 until Equal to zero or j = J, get the updated neural network structure search space S A The multi-sensor information fusion network structure parameter α A ={α O ,α P }, that is, the optimal multi-sensor fusion network structure S is obtained A* .

5. The method for multi-sensor information fusion fault diagnosis of rolling bearings as claimed in claim 4, characterized in that: Step 3.5 specifically includes the following steps: Step 3.5.1: Take the i-th sample (X trn ,Y trn ) i middle The mth sensor information Input the corresponding encoder Perform feature transformation to obtain encoded features Step 3.5.2: Encoded features Input to the fusion strategy search space Fusion and output fusion features Step 3.5.3: The fusion features obtained in step 3.5.2 Input to rolling bearing classifier h(·), output diagnosis result And calculate the diagnosis result Y i 'and The cross entropy loss Step 3.5.4, using cross entropy loss The network weight parameter w is updated to obtain the updated weight parameter w.

6. The method for multi-sensor information fusion fault diagnosis of rolling bearings as claimed in claim 5, characterized in that: Step 3.5.2 specifically includes the following steps: Step 3.5.2.1, initialize k=1; Step 3.5.2.2, determine if k≤2, if yes, proceed to step 3.5.2.3, otherwise proceed to step 3.5.2.6; Step 3.5.2.3, determine if k=1, if yes, proceed to step 3.5.2.4, otherwise proceed to step 3.5.2.5; Step 3.5.2.4, the encoded features are obtained The elements in are concatenated and copied into two identical copies as the k-th layer fusion module The first input and the second input Right now Execute step 3.5.2.7; in: Represents a splicing operation; represents the output of the k-2th layer fusion module; Step 3.5.2.5, the encoded features are obtained The elements in are concatenated as a k-layer fusion module The second input Execute step 3.5.2.7; Step 3.5.2.6, merge the k-2th layer of modules and the k-1th layer fusion module Output and As the k-th layer fusion module For the first and second inputs, execute step 3.5.2.7; Step 3.5.2.7, search space through feature extraction unit For the k-th layer fusion module The first input and the second input Perform feature extraction to obtain feature set in: Indicates that the i-th sample is fused in the k-th layer module In the example, the sensor signal of the mth sensor passes through the corresponding mth feature extraction unit Cell m Output of (N,O); and Respectively represent the information of the mth sensor of the i-th sample in the output of the k-2-layer fusion module and the k-1-layer fusion module; Step 3.5.2.8, the feature set obtained in step 3.5.2.7 Concatenate the elements in to get Where R represents the real number domain, r and d both represent the feature dimension, and r = M × l, l is the length of the feature map corresponding to each sensor signal; Step 3.5.2.9, calculate the attention score matrix A without softmax operation by the following formula o ; in: Represents the intermediate process of calculating the self-attention score; Q is the query matrix; K is the bond matrix; Step 3.5.2.10, for the attention score matrix A o Superposition mask matrix M o ∈R r×d Get the masked attention score A m =A o +M o ; The mask matrix M o There are two states: if k ≥ k f , that is, fusion is performed, then M o is a matrix of all 0s, that is, M o =0; if k <k f , that is, no fusion is performed, then M o The corresponding position elements where different information sources interact are -∞, and the rest are 0, that is, in: k f Indicates the fusion starting layer, Step 3.5.2.11, use the masked attention score A m Perform calculations on the self-attention mechanism and get the output Among them, softmax(·) represents the softmax function; V is the value matrix; Step 3.5.2.12, let k = k + 1, return to step 3.5.2.2, until k = K, output fusion features 7. A system for multi-sensor information fusion fault diagnosis of rolling bearings, based on the method for multi-sensor information fusion fault diagnosis of rolling bearings according to any one of claims 1 to 6, comprising a data acquisition unit, a model building unit and a fault detection unit; The data acquisition unit is used to simultaneously and continuously collect sensor signals of M sensors, and obtain (I+J) samples according to time periods, which are divided into training sets and validation set in: represents the i-th sample in the training set, each of which includes the sensor signals and their labels of M sensors in a time period; represents the jth sample in the validation set, each of which includes the sensor signals and their labels of M sensors in a time period; is the sensor signal of the M sensors in the i-th sample in the training set, m is the serial number of the sensor is the sensor signal of the M sensors in the jth sample in the validation set, m is the serial number of the sensor; is the sensor signal of the mth sensor in the ith sample; is the sensor signal of the mth sensor in the jth sample; is the label of the sensor signal of the M sensors in the i-th sample, and is the bearing fault category; is the label of the sensor signal of the M sensors in the jth sample, which is the bearing fault category; The model building unit is used to build a multi-sensor information fusion network structure search space S A =S O ×S P and a search space S for controlling the multi-sensor information fusion network structure A The multi-sensor information fusion network structure parameter α A ={α O ,α P }, construct rolling bearing classifier h(·); where S O represents the feature extraction unit search space, S P represents the fusion strategy search space, α O Represents the search space S used to control the feature extraction unit O The structural parameter, α P Represents the search space S used to control the fusion strategy P Structural parameters of the training set and validation set At the same time, the neural network structure search space S obtained in step 2 A The multi-sensor information fusion network structure parameter α A ={α O ,α P } and weight parameter w are updated to obtain the optimal multi-sensor fusion network structure S A* ; in: Represents the network structure parameter α from the multi-sensor information fusion network A ={α O ,α P } feature extraction unit search space S A Use search strategies in Extract a possible feature extraction unit network structure S; w represents the weight parameter of the possible feature extraction unit network structure S; To use the validation set D val The classification cross entropy loss of To use the training set D trn The classification cross entropy loss of The fault detection unit is used to collect multi-sensor signals in the scene to be diagnosed and input them into the optimal multi-sensor fusion network structure S A* In the experiment, the rolling bearing fault diagnosis results are obtained.

Images