An engineering vehicle motor fault diagnosis method and system based on a TAEB model

Abstract

The application discloses an engineering vehicle motor fault diagnosis method and system based on a TAEB model, and relates to the field of engineering vehicle motor fault diagnosis.The engineering vehicle motor fault diagnosis method comprises the following steps: constructing a motor fault data set; constructing a TAEB model for feature classification, wherein the TAEB model comprises a feature extraction network, a feature purification module, a feature fusion module and a classification and prediction module; training the TAEB model by using the motor fault data set; optimizing the trained TAEB model by using a light-weight knowledge distillation technology, and deploying the optimized TAEB model on an embedded device of an engineering vehicle; performing engineering vehicle motor fault prediction by using the optimized TAEB model, comparing the prediction result of the optimized TAEB model with a real value measured by a sensor, and obtaining an engineering vehicle motor fault diagnosis result.The application can discover potential problems of a mine car in a timely manner, facilitate an operator to take corresponding remedial measures in advance, avoid the occurrence of faults, and guarantee the safety of the engineering vehicle.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motor fault diagnosis, and particularly relates to an engineering vehicle motor fault diagnosis method and system based on a TAEB model. BACKGROUND

[0002] With the development of industrial automation and intelligentization in recent years, it has become an issue that cannot be ignored to ensure the use of electric engineering vehicles. The traditional engineering machinery equipment of the industry often needs to be manually maintained and inspected regularly, which is not efficient and is easy to overlook some problems. And in some special circumstances, due to external impact, the synchronous motor of the electric engineering vehicle is prone to accidents. The current traditional method cannot monitor the health status of the motor in real time, and lacks the early warning ability of the engineering vehicle. The electrical system of the electric mine car is complex, including the battery management system, the motor controller, the charging system, etc. Once a fault occurs, it is difficult to maintain, and the technical level of the maintenance personnel is high. For example, the maintenance and replacement cost of some key components is high, the maintenance cost of the motor controller may account for 20%-30% of the total vehicle maintenance cost, and the maintenance time is long, which may cause the vehicle to be out of service for a long time. SUMMARY

[0003] The present application provides an engineering vehicle motor fault diagnosis method and system based on a TAEB model to solve the technical problems mentioned in the background.

[0004] To achieve the above purpose, the technical scheme of the present application is as follows:

[0005] The present application provides an engineering vehicle motor fault diagnosis method based on a TAEB model, comprising the following steps:

[0006] S1, a plurality of different sensors are used to collect a plurality of different dimension running state data of the engineering vehicle, and the running state data is processed to obtain a motor fault data set;

[0007] S2, a TAEB model for feature classification is constructed, the TAEB model comprises a feature extraction network, a feature purification module, a feature fusion module and a classification prediction module, wherein the feature extraction network comprises a feature extraction module based on a Transformer model and a feature extraction module based on a BiLSTM model connected in parallel; the feature extraction module based on the Transformer model is connected with the feature purification module, and the output ends of the feature purification module and the feature extraction module based on the BiLSTM model are connected with the input end of the feature fusion module; the output end of the feature fusion module is connected with the classification prediction module;

[0008] S3, the TAEB model is trained using the motor fault data set to obtain a trained TAEB model;

[0009] S4, optimizing the trained TAEB model by using a light-weight knowledge distillation technology, and quantizing and deploying the optimized TAEB model on an embedded device of the engineering vehicle;

[0010] S5, using the quantized TAEB model on the embedded device to perform engineering vehicle motor fault prediction, and then comparing the prediction result of the quantized TAEB model with the real value measured by the sensor to obtain an engineering vehicle motor fault diagnosis result.

[0011] Further, the S1 specifically comprises the following steps:

[0012] S11, collecting a plurality of different dimension running state data of the engineering vehicle by using a plurality of types of sensors on the engineering vehicle, the running state data including current, voltage and speed of the vehicle motor;

[0013] S12, preprocessing and labeling the plurality of different dimension running state data, and then integrating the labeled running state data by using a data fusion algorithm to obtain integrated running state data;

[0014] S13, converting and normalizing the integrated running state data to eliminate dimension differences between features, to obtain a motor fault data set; and then dividing the motor fault data set into a training set and a test set.

[0015] Further, the feature extraction module based on the Transformer model comprises a position encoder and a plurality of layers of Transformer encoders connected in series, for layer-by-layer extraction of features, and outputs final multi-scale deep features;

[0016] The feature extraction module based on the BiLSTM model comprises a forward LSTM unit and a backward LSTM unit connected to each other, and the forward LSTM unit and the backward LSTM unit are connected to the feature fusion module through a summation operation, wherein the forward LSTM unit comprises a plurality of long short-term memory networks LSTM R connected in series, and the backward LSTM unit comprises a plurality of long short-term memory networks LSTM L ;

[0017] The feature purification module comprises a self-encoder MEQAE and a decoder connected to each other.

[0018] The feature fusion module comprises a bidirectional long short-term memory network, a self-attention mechanism, an attention fusion layer and a fully connected layer connected in series.

[0019] Further, the S3 specifically comprises the following steps:

[0020] S31, input the original input data in the training set into the feature extraction module based on the Transformer model, and output the final multi-scale deep features;

[0021] S32, input the original input data of the training set into the feature extraction module based on the BiLSTM model, and obtain the time sequence features;

[0022] S33, input the final multi-scale deep features into the autoencoder MEQAE of the feature purification module, obtain the purified feature representation, input the purified feature representation into the decoder to obtain the reconstruction calculation result, and input the purified feature representation as the refined feature representation into the feature fusion module together with the time sequence features;

[0023] S34, after the feature fusion module receives the refined feature representation and the time sequence features, output the comprehensive feature vector;

[0024] S35, input the comprehensive feature vector and the purified feature representation in S33 into the classification prediction module to obtain the classification diagnosis result with explicit confidence;

[0025] S36, construct a loss function according to the classification diagnosis result with explicit confidence and the pre-labeled real label, and adjust the TAEB model through the loss function;

[0026] S37, determine whether the preset iteration stop condition is reached, if yes, output the TAEB model after parameter adjustment, otherwise, cycle S31 to S36 until the preset iteration stop condition is reached, and output the TAEB model after parameter adjustment;

[0027] S38, test the TAEB model after parameter adjustment by using the test set to obtain the test result, determine whether the test result reaches the preset deployment standard, if yes, output the trained TAEB model, otherwise, return to S31 to train again until the preset deployment standard is reached, and output the trained TAEB model.

[0028] Further, the S31 specifically includes the following steps:

[0029] S311, input the training set into the feature extraction module based on the Transformer model, first map the original input data in the training set into initial embedding vectors through linear projection, and inject sine cosine position encoding through the position encoder to generate position-aware embedding sequence, which is expressed by the formula as follows:

[0030] ;

[0031] ;

[0032] in, This represents the output of the sinusoidal position encoding; This indicates the output of the cosine position encoding; Indicates the position index of an element in the embedded sequence; This represents the feature dimension of the feature extraction module in the Transformer model. The dimension index represents the positional encoding vector. The range of values ​​is ; This represents the control coefficient, used to control the periodicity of position information;

[0033] S312. The position-aware embedding sequence is input into the multi-head self-attention mechanism in the first layer Transformer encoder. The multi-head self-attention mechanism calculates the correlation between elements within the sequence based on the position-aware embedding sequence, dynamically generates global context association weights, and uses the global context association weights to aggregate sequence information, finally outputting a context-aware feature sequence.

[0034] S313. Perform residual connection and layer normalization on the context-aware feature sequence and the position-aware embedding sequence, and then perform nonlinear transformation on the output result through the feedforward neural network to obtain the transformed features.

[0035] S314. The transformed features are again connected by residuals and normalized by layers to form the output of the first layer Transformer encoder.

[0036] S315. Input the output of the first layer Transformer encoder into the subsequent layers of Transformer encoders to output the final multi-scale depth features.

[0037] Furthermore, step S33 specifically includes the following steps:

[0038] S331. The final multi-scale deep features are input into the feature purification module. The autoencoder MEQAE maps the final multi-scale deep features to a latent space of a first set dimension and generates a latent encoding vector in this latent space. In the latent encoding vector Two types of regularization constraints are applied: one is a sparsity constraint, which is applied to the loss function. Add latent encoding vector The L1 norm is used as a penalty to drive the latent encoding vector. Sparsity enables feature selection; secondly, it addresses bottleneck constraints, reducing the potential encoding vector. The dimension of the latent encoding vector is lower than the dimension of the final multi-scale deep features, forcing the latent encoding vector to... information compression is performed; after the regularization constraint, a purified feature representation is obtained;

[0039] The expression of the autoencoder MEQAE is as follows:

[0040] ;

[0041] wherein, denotes the hidden state of the long short-term memory network LSTM at time step t; denotes an h-dimensional real vector; h denotes the dimension of the hidden state; denotes the t-th time step of the time series input; denotes an n-dimensional real vector; n denotes the dimension of the input data; denotes the set of real numbers; denotes the latent encoding vector by the mapping function converts the hidden state at the last time step into the latent encoding vector

[0042] The bottleneck constraint is that the deep features above a first set dimension in the latent encoding vector are forced to be compressed by the autoencoder MEQAE, so as to force the feature purification module to filter noise and redundant information;

[0043] The bottleneck constraint is specifically expressed by the formula as follows:

[0044] ;

[0045] wherein, denotes the reconstruction loss function, and the reconstruction loss function is calculated by the reconstruction calculation result; denotes the autoencoder MEQAE; denotes the decoder; denotes being limited to; denotes the dimension of the latent encoding vector is less than the dimension of the final multi-scale deep feature X;

[0046] The sparse constraint is that the L1 norm regularization is applied to the latent encoding vector by the autoencoder MEQAE as a penalty, so as to force the activation values of the dimensions above a second set proportion in the latent encoding vector to be suppressed to zero or close to zero, and only allow the features of the dimensions below the second set proportion to remain activated;

[0047] The expression of the sparse constraint is specifically as follows:

[0048] ;

[0049] wherein, represents a loss function; represents the number of KL divergence terms; represents a KL divergence term; represents a weight hyper-parameter of the KL divergence regular term; represents a prior probability; represents a probability estimation output by the feature purification module; represents a weight hyper-parameter of the sparsity constraint term, which is used to control the intensity of sparsity of the latent encoding vector ; the weight hyper-parameter is larger, the stronger the constraint on sparsity is; KL represents a KL divergence loss; represents the norm of the latent encoding vector ;

[0050] S332, input the purified feature representation into the decoder to perform internal reconstruction calculation to obtain a reconstruction calculation result, and simultaneously, the purified feature representation is taken as a refined feature representation output by the feature purification module and is input into the feature fusion module;

[0051] The expression of the decoder is specifically as follows:

[0052] ;

[0053] ;

[0054] wherein, represents a hidden state predicted by the feature purification module at a time step t; represents an input of the feature purification module at a time step t; represents a final prediction output of the feature purification module at a time step t.

[0055] Further, the S34 specifically comprises the following steps:

[0056] S341, after the feature fusion module receives the refined feature representation and the time sequence feature, the feature fusion module internally uses a bidirectional long short-term memory network to perform time sequence modeling on the refined feature to capture the front-back dependency relationship of the refined feature, and simultaneously, uses a self-attention mechanism to enhance the internal correlation of the global feature to obtain an enhanced global feature;

[0057] The expression of the bidirectional long short-term memory network is as follows:

[0058] ;

[0059] ;

[0060] wherein, denotes the cell state of the bidirectional long short-term memory network at time step t; denotes the output of the forget gate in the bidirectional long short-term memory network at time step t; denotes the candidate cell state at time step t; denotes the output of the input gate in the bidirectional long short-term memory network at time step t; denotes element-wise multiplication; denotes the output of the output gate in the bidirectional long short-term memory network;

[0061] S342, concatenate the enhanced global feature and the time sequence feature, and pass them through a learnable attention fusion layer to perform dynamic weight distribution and adaptive weighted fusion of local dynamics and global context to obtain a fused feature;

[0062] S343, input the fused feature into a fully connected layer for nonlinear transformation and Dropout regularization processing, and output a comprehensive feature vector.

[0063] Further, the S4 specifically comprises the following steps:

[0064] S41, lightweight design of the trained TAEB model, including: reducing the size and number of layers of the hidden layer of the long short-term memory network and the bidirectional long short-term memory network, and simplifying the structure and dimension of the fully connected layer in the classification prediction module to obtain a lightweight designed TAEB model;

[0065] S42, taking the lightweight designed TAEB model as a student model, and taking the trained TAEB model without lightweight design as a teacher model;

[0066] S43, training the student model, and taking the output of the teacher model as supervision information in each training to guide the student model to learn, and constructing a total loss function in each training to dynamically adjust the student model, and obtaining an optimized TAEB model after training;

[0067] S44, quantizing the optimized TAEB model and deploying it on the embedded device of the engineering vehicle.

[0068] Further, the total loss function in the S43 is specifically as follows:

[0069] ;

[0070] wherein, denotes the total loss function; denotes the cross-entropy loss; denotes the knowledge distillation loss; denotes a regularization coefficient for balancing the weights of the cross-entropy loss and the knowledge distillation loss;

[0071] The calculation formula of the knowledge distillation loss is as follows:

[0072]

[0073] wherein, denotes a target distribution; denotes a predicted distribution; and are output probabilities of the teacher model and the student model in the i-th class, respectively, denotes a summation calculation over all fault classes. Another aspect of the present application also provides an engineering vehicle motor fault diagnosis system configured or performing the above engineering vehicle motor fault diagnosis method.

[0074] The present application has the following beneficial effects:

[0075] The method provided by the present application can enable the driver or safety officer to discover potential problems of the engineering vehicle motor in time, realize fault diagnosis and early warning of the engineering vehicle motor, so that the driver or safety officer can take corresponding remedial measures in advance, thereby avoiding sudden occurrence or aggravation of the engineering vehicle motor fault as much as possible, and ensuring safe and stable operation of the engineering vehicle.

[0076] In addition, the present application realizes fault diagnosis of the engineering vehicle motor, provides a motor repair basis for motor repair personnel, saves the time spent by the motor repair personnel in detecting the faulty equipment, improves the repair efficiency, and greatly facilitates the motor repair personnel.

[0077] In addition, the present application adopts the form of cooperation of multiple sensors and an optimized TAEB model for engineering vehicle motor fault diagnosis, and compared with the traditional fault diagnosis system, the structure is simpler, the equipment cost is lower, and it is also conducive to later maintenance and maintenance.

[0078] In addition, the present application adopts the form of cooperation of multiple sensors and an optimized TAEB model for engineering vehicle motor fault diagnosis, and compared with the traditional fault diagnosis system, the structure is simpler, the equipment cost is lower, and it is also conducive to later maintenance and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 is a structural block diagram of the TAEB model in the present application;

[0080] Figure 2 is a structural block diagram of the feature extraction module based on the Transformer model in the present application;

[0081] Figure 3 is a structural block diagram of the feature extraction module based on the BiLSTM model in the present application;​​​

[0082] Figure 4 Flow chart of the method for diagnosing faults of an electric machine of an engineering vehicle according to the application. DETAILED DESCRIPTION

[0083] For the purpose of promoting the understanding of the present application, a more complete description will be made below with reference to the accompanying drawings. The preferred embodiments of the present application are shown in the drawings. However, the present application can be realized in many other different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided in order to make the disclosure of the present application more thorough and comprehensive.

[0084] In addition, the terms "first", "second" are only used for the purpose of description and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0085] With reference to Figure 1 and Figure 4 The embodiment of the present application provides a method for diagnosing faults of an electric machine of an engineering vehicle based on a TAEB model, comprising the following steps:

[0086] S1, collecting running state data of the engineering vehicle in multiple different dimensions by using multiple different sensors, and processing the running state data to obtain an electric machine fault data set;

[0087] S2, constructing a TAEB model for feature classification, the TAEB model comprising a feature extraction network, a feature purification module, a feature fusion module and a classification prediction module, wherein the feature extraction network comprises a feature extraction module based on a Transformer model and a feature extraction module based on a BiLSTM model connected in parallel; the feature extraction module based on the Transformer model is connected with the feature purification module, and the output ends of the feature purification module and the feature extraction module based on the BiLSTM model are connected with the input end of the feature fusion module; the output end of the feature fusion module is connected with the classification prediction module;

[0088] S3, training the TAEB model by using the electric machine fault data set to obtain a trained TAEB model;

[0089] S4, optimizing the trained TAEB model by using a light-weight knowledge distillation technology, and deploying the optimized TAEB model on an embedded device of the engineering vehicle after quantization;

[0090] S5, using the quantized TAEB model on the embedded device to perform engineering vehicle motor fault prediction, and then comparing the prediction result of the quantized TAEB model with the true value measured by the sensor to obtain the engineering vehicle motor fault diagnosis result.

[0091] The method provided by the application can enable the driver or the safety officer to find the potential problem of the engineering vehicle motor in time, realize fault diagnosis and early warning of the engineering vehicle motor, so that the driver or the safety officer can take corresponding maintenance measures in advance, thereby avoiding sudden occurrence or aggravation of the engineering vehicle motor fault as much as possible, and ensuring safe and stable operation of the engineering vehicle.

[0092] In addition, the application realizes fault diagnosis of the engineering vehicle motor, provides a motor repair basis for motor repair personnel, saves the time spent by the motor repair personnel in detecting the faulty equipment, improves the repair efficiency, and greatly facilitates the motor repair personnel.

[0093] In addition, the application adopts the form of cooperation of multiple sensors and an optimized TAEB model to perform engineering vehicle motor fault diagnosis, and compared with a traditional fault diagnosis system, the structure is simpler, the equipment cost is lower, and the post-maintenance and maintenance are also facilitated.

[0094] In some embodiments, the S1 specifically comprises the following steps:

[0095] S11, collecting multiple types of running state data of the engineering vehicle in multiple dimensions by using multiple types of sensors (including current sensors, voltage sensors, speed sensors, etc.) on the engineering vehicle, wherein the running state data includes the current, voltage and speed of the vehicle motor;

[0096] S12, preprocessing and labeling the running state data in multiple dimensions, and then integrating and processing the labeled running state data by using a data fusion algorithm to obtain integrated running state data, wherein the preprocessing includes uniform coding, cleaning and denoising, etc., wherein the uniform coding is to convert the original data (i.e., the running state data) from different sources and different formats into a machine-readable and format-unified numerical representation form to ensure the consistency and processability of the data;

[0097] S13, converting and normalizing the integrated running state data to eliminate the dimensional difference between the features to obtain a motor fault data set, and then dividing the motor fault data set into a training set and a test set according to a set proportion (such as 8:2). After the division, it is checked whether there are residual missing values in the training set and the test set, whether the data statistical characteristics are reasonable, and whether the data distribution meets the expectation, etc.

[0098] The expression of the normalization processing in the above is as follows:

[0099] ;

[0100] ;

[0101] wherein, denotes a vector to be normalized; denotes a normalized vector; denotes a p-norm; denotes a certain dimension component of.

[0102] In this embodiment, the value of p is 2, that is, L2 normalization is adopted, and the dimension difference between features can be eliminated through L2 normalization;

[0103] In some embodiments, referring to Figure 2 , the feature extraction module based on the Transformer model comprises a position encoder and a multi-layer Transformer encoder connected in sequence, for layer-by-layer extraction of the features, and outputs the final multi-scale deep features;

[0104] Referring to Figure 3 , the feature extraction module based on the BiLSTM model comprises a forward LSTM unit (or left LSTM unit) and a backward LSTM unit (or right LSTM unit) connected to each other, and the forward LSTM unit and the backward LSTM unit are connected to the feature fusion module through a summation operation, wherein the forward LSTM unit comprises a plurality of long short-term memory networks LSTM R responsible for processing the forward timing information of the sequence, and the backward LSTM unit comprises a plurality of long short-term memory networks LSTM L responsible for processing the reverse timing information of the sequence.

[0105] The feature purification module comprises a self-encoder MEQAE and a decoder connected to each other.

[0106] The feature fusion module comprises a bidirectional long short-term memory network, a self-attention mechanism, an attention fusion layer and a fully connected layer connected in sequence.

[0107] The present application discloses a TAEB model, wherein the letter T in TAEB represents the Transformer model, which corresponds to the feature extraction module based on the Transformer model in the present application, the letter AE represents the self-encoder MEQAE, and the letter B represents the BiLSTM model, which corresponds to the feature extraction module based on the BiLSTM model in the present application.

[0108] In some embodiments, the S3 specifically comprises the following steps:

[0109] S31, input the original input data in the training set into the feature extraction module based on the Transformer model, and output the final multi-scale deep features;

[0110] S32, input the original input data of the training set into the feature extraction module based on the BiLSTM model, and obtain the time sequence feature;

[0111] S33, input the final multi-scale deep features into the autoencoder MEQAE of the feature purification module, obtain the purified feature representation, input the purified feature representation into the decoder to obtain the reconstruction calculation result, and input the purified feature representation as the refined feature representation into the feature fusion module together with the time sequence feature;

[0112] S34, after the feature fusion module receives the refined feature representation and the time sequence feature, output the comprehensive feature vector;

[0113] S35, input the comprehensive feature vector and the purified feature representation in S33 into the classification prediction module to obtain the classification diagnosis result with explicit confidence;

[0114] S36, construct a loss function according to the classification diagnosis result with explicit confidence and the pre-labeled real label, and adjust the TAEB model through the loss function;

[0115] S37, determine whether the preset iteration stopping condition is reached, if yes, output the TAEB model after parameter adjustment, otherwise, cycle S31 to S36 until the preset iteration stopping condition is reached, and output the TAEB model after parameter adjustment;

[0116] S38, test the TAEB model after parameter adjustment by using the test set to obtain a test result, determine whether the test result reaches the preset deployment standard, if yes, output the trained TAEB model, otherwise, return to S31 to train again until the preset deployment standard is reached, and output the trained TAEB model.

[0117] In some embodiments, the S31 specifically includes the following steps:

[0118] S311, input the training set into the feature extraction module based on the Transformer model, first map the original input data in the training set into initial embedding vectors through linear projection, in order to make up for the defect that the Transformer model is not sensitive to sequence order, inject sine cosine position encoding through a position encoder, so that the feature extraction module based on the Transformer model can perceive the relative position of elements, generate a position-aware embedding sequence, and the formula is as follows:

[0119] ;

[0120] ;

[0121] wherein, denotes the output of the sine position encoding; denotes the output of the cosine position encoding; denotes the position index of the element in the embedding sequence; denotes the feature dimension of the feature extraction module of the Transformer model; denotes the dimension index of the position encoding vector, the dimension index has a value range of ; denotes a control coefficient for controlling the periodicity of the position information; in the embodiment, the control coefficient has a value of 10000;

[0122] S312, input the position-aware embedding sequence into the multi-head self-attention mechanism in the first layer of the Transformer encoder, and calculate the correlation degree between the elements in the sequence based on the position-aware embedding sequence, dynamically generate global context correlation weights, and aggregate the sequence information using the global context correlation weights, and finally output a context-aware feature sequence;

[0123] The expression of the single-head attention mechanism is specifically as follows:

[0124] ;

[0125] wherein, denotes the output of the single-head attention mechanism; denotes an activation function, denote a query vector, a key vector, and a value vector, respectively; denotes the dimension size of ; T denotes transposition;

[0126] S313, perform residual connection and layer normalization on the context-aware feature sequence and the position-aware embedding sequence, and then perform nonlinear transformation on the output result through a feedforward neural network to obtain a transformed feature;

[0127] S314, the transformed feature is again connected in residual connection and normalized in layers to form the output of the first layer of the Transformer encoder;

[0128] S315, input the output of the first layer of the Transformer encoder into the subsequent layers of the Transformer encoder, and output the final multi-scale deep feature.

[0129] In some embodiments, the S33 specifically comprises the following steps:

[0130] S331, input the final multi-scale deep feature into a feature purification module, and the autoencoder MEQAE maps the final multi-scale deep feature to a latent space with a first set dimension below, and generates a latent encoding vector in the latent space , and the latent encoding vector is subjected to two regularization constraints: one is a sparse constraint, and the L1 norm of the latent encoding vector is added as a penalty in the loss function to drive the latent encoding vector to be sparse, realizing feature selection; the other is a bottleneck constraint, which makes the dimension of the latent encoding vector lower than the dimension of the final multi-scale deep feature, forcing the latent encoding vector to compress information; after the regularization constraint, a purified feature representation is obtained;

[0131] The expression of the final multi-scale deep feature is:

[0132] Wherein ;

[0133] Wherein, represents the final multi-scale deep feature, represents the deep feature of different scales;

[0134] The expression of the autoencoder MEQAE is as follows:

[0135] ;

[0136] Wherein, represents the hidden state of the long short-term memory network LSTM at time step t; represents an h-dimensional real vector; h represents the dimension of the hidden state; represents the t-th time step of the time series input; represents an n-dimensional real vector; n represents the dimension of the input data; represents the set of real numbers; represents the hidden state at the last time step converted into the latent encoding vector by the mapping function ;

[0137] The bottleneck constraint is that the latent encoding vector The deep features above a first set of dimensions are forced to be compressed, so that the feature purification module is forced to filter noise and redundant information;

[0138] The formula of the bottleneck constraint is as follows:

[0139] ;

[0140] wherein, represents a reconstruction loss function, and the reconstruction loss function is calculated by the reconstruction calculation result; represents the autoencoder MEQAE; represents the decoder; is limited to represents the latent encoding vector The dimension of the latent encoding vector is far less than the dimension of the final multi-scale deep feature X.

[0141] The sparse constraint is that the latent encoding vector is subjected to L1 norm regularization as a penalty, so that the latent encoding vector The activation values of more than a first set of proportions (such as the majority) of the dimensions are suppressed to zero or close to zero, and only the features of less than a second set of proportions (such as the minority) of the dimensions are allowed to remain active; this mechanism drives the feature purification module to automatically perform feature selection, suppresses redundant information and unnecessary features, and thus learns a highly simplified and discriminative sparse feature representation;

[0142] The expression of the sparse constraint is as follows:

[0143] ;

[0144] wherein, represents a loss function; represents the number of KL divergence terms; represents a KL divergence term; represents a weight hyperparameter of the KL divergence regularization term; represents a prior probability; represents a probability estimate output by the feature purification module; represents a weight hyperparameter of the sparse constraint term, used to control the strength of the sparsity of the latent encoding vector The greater the weight hyperparameter , the stronger the constraint on sparsity; KL represents the KL divergence loss; represents the norm of the latent encoding vector ;

[0145] S332, input the purified feature representation into the decoder, enhance the robustness by injecting noise combined with quantum anti-interference characteristics, and promote feature decoupling with quantum state orthogonalization to realize internal reconstruction calculation, obtain a reconstruction calculation result, the reconstruction calculation result is a refined feature with high discriminability, and the purified feature representation is also output as a refined feature representation of the feature purification module and input into the feature fusion module; in addition, the purified feature representation is directly fed to the classification prediction module;

[0146] The expression of the decoder is specifically as follows:

[0147] ;

[0148] ;

[0149] wherein, represents the hidden state of the feature purification module predicted at time step t; represents the input of the feature purification module at time step t; represents the final prediction output of the feature purification module at time step t.

[0150] In some embodiments, the S34 specifically comprises the following steps:

[0151] S341, after the feature fusion module receives the refined feature representation and the time sequence feature, the bidirectional long short-term memory network inside the feature fusion module is used to model the time sequence of the refined feature to capture the before-and-after dependency of the refined feature, and the self-attention mechanism is used to enhance the internal association of the global feature, to obtain an enhanced global feature;

[0152] The expression of the bidirectional long short-term memory network is as follows:

[0153] ;

[0154] ;

[0155] wherein, represents the cell state of the bidirectional long short-term memory network at time step t; represents the output of the forget gate in the bidirectional long short-term memory network at time step t; represents the candidate cell state at time step t; represents the output of the input gate in the bidirectional long short-term memory network at time step t; represents element-wise multiplication; represents the output of the output gate in the bidirectional long short-term memory network;

[0156] S342, the enhanced global feature is spliced with the time sequence feature, and a learnable attention fusion layer is used for dynamic weight distribution, and adaptive weighted fusion of local dynamics and global context is performed to obtain a fused feature;

[0157] Specifically, the feature fusion module performs weighted fusion on the enhanced global feature and the time sequence feature, dynamically adjusts the contribution degrees of the two features through a learnable attention fusion layer, and the calculation formulas are as follows:

[0158] ;

[0159] ;

[0160] wherein, represents a weight vector used for privilege weighting finally; represents an activation function; represents a transformation matrix; represents an enhanced global feature; represents a time sequence feature; represents a feature splicing operation; represents a fused feature; 、 represents a fusion coefficient; the application balances the contributions of the enhanced global feature and the time sequence feature, and avoids the limitations of a single mechanism;

[0161] S343, the fused feature is sent to a fully connected layer for nonlinear transformation and Dropout regularization processing, and an integrated feature vector is output.

[0162] In some embodiments, the S35 specifically includes the following steps:

[0163] S351, the classification prediction module receives the integrated feature vector output by the feature fusion module, first maps the integrated feature vector to a Logits vector with the same dimension and the same number of fault categories through one or more fully connected layers;

[0164] S352, then, the Logits vector is input to a Softmax function for normalization processing, and the Logits vector is converted into a probability distribution with a sum of all elements being 1;

[0165] S353, in this probability distribution, the probability value corresponding to the “healthy” category is defined as the “health probability”, which is used as the quantitative confidence for evaluating the health state of the system;

[0166] S354, the potential coding is time-series modeled to capture the before-after dependency of the sequence, and is mapped to the category probability of the motor fault through a full connection layer and a Dropout operation (to randomly set the neuron output to zero with a certain probability);

[0167] The expression of the full connection layer is specifically as follows:

[0168]

[0169] wherein, represents the output of the full connection layer, represents a weight coefficient; represents the input of the full connection layer; represents a bias parameter;

[0170] In some embodiments, the S4 specifically comprises the following steps:

[0171] S41, the trained TAEB model is designed to be lightweight, including: reducing the size and number of hidden layers of the long short-term memory network and the bidirectional long short-term memory network, and simplifying the structure and dimension of the full connection layer in the classification prediction module, to obtain a lightweight designed TAEB model;

[0172] S42, the lightweight designed TAEB model is taken as a student model, and the trained TAEB model without lightweight design is taken as a teacher model; the teacher model usually has a higher accuracy, but may have a higher calculation cost or a complex structure. The student model is relatively simple, and aims to improve the performance by learning the knowledge of the teacher model;

[0173] S43, the student model is trained, and the output of the teacher model is taken as supervision information in each training to guide the student model to learn, so that the student model can obtain richer semantic information from the teacher model, thereby achieving good performance under a smaller model size. In each training, a total loss function is constructed, and the student model is dynamically adjusted according to the total loss function. After training, an optimized TAEB model is obtained;

[0174] S44, the optimized TAEB model is quantized and deployed on the embedded device of the engineering vehicle.

[0175] Specifically, the optimized TAEB model is applied to dynamic quantization to convert 32-bit floating-point parameters to 8-bit integers, further reducing the model size and improving the inference speed. Then, the quantized TAEB model is converted into an ONNX format, so as to facilitate the use of a lightweight inference engine ONNX Runtime for deployment on the embedded device of the engineering vehicle.

[0176] In some embodiments, the total loss function in the S43 is specifically as follows:​

[0177] ;

[0178] in, Represents the total loss function; Represents cross-entropy loss; This represents the loss from knowledge distillation. This represents the regularization coefficient, used to balance the weights of cross-entropy loss and knowledge distillation loss; by adjusting the weights... The training of student models can be optimized based on specific tasks and model characteristics;

[0179] In the formula, knowledge distillation loss The calculation formula is as follows:

[0180] ;

[0181] in, Indicates the target distribution; Indicates the predicted distribution; and The teacher model and the student model are respectively in the 1st month. Output probabilities for each category This indicates all fault categories. Perform a summation calculation.

[0182] In some embodiments, S5 specifically includes the following steps:

[0183] S51. The quantized TAEB model is used to monitor the motor of the engineering vehicle. Then, the prediction results of the quantized TAEB model are compared with the real values ​​collected in real time by multiple sensors, and the residuals are obtained. The formula is expressed as follows:

[0184] ;

[0185] in, Represents the residual; Represents the first output of multiple transmitters The true value of each sample; This represents the prediction results of the quantized TAEB model;

[0186] S52, judge whether the residual error is in the set range fluctuation, if yes, it indicates that the quantized TAEB model is relatively accurate for predicting the running state of the motor of the engineering vehicle, the motor of the engineering vehicle is relatively stable, and no abnormal situation may occur. If the residual error suddenly increases or abnormally fluctuates, it indicates that the difference between the predicted value and the true value of the quantized TAEB model is large, which may mean that the motor of the engineering vehicle has a fault or the running state has an abnormal change. For example, in terms of the motor current of the engineering vehicle, the too large residual error may imply that the motor of the engineering vehicle has a short circuit, an open circuit or the like.

[0187] Through continuous monitoring and analysis of the residual error, the safety officer or operator can find the potential problems of the motor of the engineering vehicle in time, take corresponding remedial measures in advance, avoid the occurrence or aggravation of the fault, and ensure the safe and stable operation of the motor of the engineering vehicle.

[0188] Another aspect of the present application also provides an engineering vehicle motor fault diagnosis system configured or performing the above-mentioned engineering vehicle motor fault diagnosis method.

[0189] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Moreover, the technical solutions of each embodiment of the present application can be combined with each other, but it must be based on the realization of the ordinary skilled person in the art. When the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist and is not within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for diagnosing motor faults in engineering vehicles based on the TAEB model, characterized in that, Includes the following steps: S1. Multiple different sensors are used to collect operating status data of engineering vehicles from multiple dimensions, and the operating status data is processed to obtain motor fault dataset; S2. Construct a TAEB model for feature classification. The TAEB model includes a feature extraction network, a feature purification module, a feature fusion module, and a classification prediction module. The feature extraction network includes a parallel feature extraction module based on a Transformer model and a feature extraction module based on a BiLSTM model. The feature extraction module based on the Transformer model is connected to the feature purification module. The outputs of both the feature purification module and the feature extraction module based on the BiLSTM model are connected to the input of the feature fusion module. The output of the feature fusion module is connected to the classification prediction module. S3. The TAEB model is trained using the motor fault dataset to obtain the trained TAEB model; S4. Optimize the trained TAEB model using lightweight knowledge distillation technology, and then quantize and deploy the optimized TAEB model on the embedded device of the engineering vehicle. S5. Use the quantized TAEB model on the embedded device to predict the motor fault of the engineering vehicle, and then compare the prediction result of the quantized TAEB model with the real value obtained by the sensor to obtain the motor fault diagnosis result of the engineering vehicle. The feature purification module includes an autoencoder MEQAE and a decoder that are interconnected. S4 specifically includes the following steps: S41. Lightweight design of the trained TAEB model, including: reducing the size and number of hidden layers of the long short-term memory network and bidirectional long short-term memory network, and simplifying the structure and dimension of the fully connected layer in the classification prediction module, to obtain a lightweight TAEB model. S42. Use the lightweight TAEB model as the student model and the TAEB model trained without lightweight design as the teacher model. S43. Train the student model and use the output of the teacher model as supervision information in each training session to guide the student model to learn. In each training session, construct the total loss function and dynamically adjust the student model based on the total loss function. After training, the optimized TAEB model is obtained. S44. Quantize the optimized TAEB model and deploy it on the embedded device of the engineering vehicle.

2. The method for diagnosing motor faults in engineering vehicles based on the TAEB model according to claim 1, characterized in that, S1 specifically includes the following steps: S11. Collect multiple types of sensors on the engineering vehicle to collect operating status data of the engineering vehicle in multiple dimensions, including the current, voltage and speed of the vehicle motor. S12. Preprocess and label the operation status data of multiple different dimensions with real labels, and then use a data fusion algorithm to integrate the labeled operation status data to obtain the integrated operation status data. S13. The integrated operating status data is transformed and normalized to eliminate the differences in the units between features, and a motor fault dataset is obtained. Then the motor fault dataset is divided into a training set and a test set.

3. The method for diagnosing motor faults in engineering vehicles based on the TAEB model according to claim 2, characterized in that, The feature extraction module based on the Transformer model includes a position encoder and a multi-layer Transformer encoder connected in series, which are used to extract features layer by layer and output the final multi-scale depth features. The feature extraction module based on the BiLSTM model includes interconnected forward LSTM units and backward LSTM units. These units are connected to the feature fusion module via a summation operation. The forward LSTM unit comprises multiple cascaded Long Short-Term Memory (LSTM) networks. R The backward LSTM unit consists of multiple Long Short-Term Memory (LSTM) networks connected in series. L ; The feature fusion module includes a bidirectional long short-term memory network, a self-attention mechanism, an attention fusion layer, and a fully connected layer connected in sequence.

4. The method for diagnosing motor faults in engineering vehicles based on the TAEB model according to claim 1, characterized in that, S3 specifically includes the following steps: S31. Input the original input data in the training set into the feature extraction module based on the Transformer model, and output the final multi-scale deep features. S32. Input the original input data of the training set into the feature extraction module based on the BiLSTM model to obtain the time series features; S33. The final multi-scale depth features are input into the autoencoder MEQAE of the feature purification module to obtain the purified feature representation. The purified feature representation is then input into the decoder to obtain the reconstruction calculation result. At the same time, the purified feature representation is used as the refined feature representation and input into the feature fusion module together with the temporal features. S34. After receiving the refined feature representation and temporal features, the feature fusion module outputs a comprehensive feature vector. S35. Input the comprehensive feature vector and the purified feature representation from S33 into the classification prediction module to obtain a classification diagnosis result with a clear confidence level. S36. Construct a loss function based on the classification diagnosis results with clear confidence and the pre-labeled true labels, and adjust the TAEB model through the loss function; S37. Determine whether the preset iteration stop condition has been met. If so, output the TAEB model with adjusted parameters. Otherwise, loop from S31 to S36 until the preset iteration stop condition is met, and output the TAEB model with adjusted parameters. S38. Test the TAEB model with adjusted parameters using the test set, obtain the test results, and determine whether the test results meet the preset deployment standard. If yes, output the trained TAEB model; otherwise, return to S31 to train again until the preset deployment standard is met, and output the trained TAEB model.

5. The method for diagnosing motor faults in engineering vehicles based on the TAEB model according to claim 4, characterized in that, S31 specifically includes the following steps: S311. The training set is input into the feature extraction module based on the Transformer model. First, the original input data in the training set is mapped to an initial embedding vector through linear projection. Then, sine and cosine position codes are injected through the position encoder to generate a position-aware embedding sequence, which is expressed by the following formula: in, This represents the output of the sinusoidal position encoding; This indicates the output of the cosine position encoding; Indicates the position index of an element in the embedded sequence; This represents the feature dimension of the feature extraction module in the Transformer model. The dimension index represents the positional encoding vector. The range of values ​​for is [0, / 2]; This represents the control coefficient, used to control the periodicity of position information; S312. The position-aware embedding sequence is input into the multi-head self-attention mechanism in the first layer Transformer encoder. The multi-head self-attention mechanism calculates the correlation between elements within the sequence based on the position-aware embedding sequence, dynamically generates global context association weights, and uses the global context association weights to aggregate sequence information, finally outputting a context-aware feature sequence. S313. Perform residual connection and layer normalization on the context-aware feature sequence and the position-aware embedding sequence, and then perform nonlinear transformation on the output result through the feedforward neural network to obtain the transformed features. S314. The transformed features are again connected by residuals and normalized by layers to form the output of the first layer Transformer encoder. S315. Input the output of the first layer Transformer encoder into the subsequent layers of Transformer encoders to output the final multi-scale depth features.

6. The method for diagnosing motor faults in engineering vehicles based on the TAEB model according to claim 5, characterized in that, S33 specifically includes the following steps: S331. The final multi-scale deep features are input into the feature purification module. The autoencoder MEQAE maps the final multi-scale deep features to a latent space of a first set dimension and generates a latent encoding vector in this latent space. In the latent encoding vector Two types of regularization constraints are applied: one is a sparsity constraint, which is applied to the loss function. Add latent encoding vector of Norms, as a penalty, drive the latent encoding vector Sparsity enables feature selection; secondly, it addresses bottleneck constraints, reducing the potential encoding vector. The dimension of the latent encoding vector is lower than the dimension of the final multi-scale deep features, forcing the latent encoding vector to... Information compression is performed; after regularization constraints, a purified feature representation is obtained. The expression for the autoencoder MEQAE is as follows: in, This indicates that the Long Short-Term Memory (LSTM) network operates at time steps. t The hidden state; express h A dimensional real vector; h The dimension representing the hidden state; Represents the first of the timing inputs t One time step; This represents an n-dimensional real vector; n represents the dimension of the input data. Represents the set of real numbers; Indicates through mapping function Hidden state at the last moment Convert to latent encoding vector ; The bottleneck constraint is: encoding the potential vector using the autoencoder MEQAE. The deep features above the first set dimension are forcibly compressed, forcing the feature purification module to filter out noise and redundant information. The bottleneck constraint is expressed by the following formula: in, Represents the reconstruction loss function. Calculated by reconstructing the calculation results; This represents the self-encoder MEQAE; Indicates decoder; Indicates that it is subject to; Represents the latent encoding vector The dimension is smaller than the final multi-scale deep features X The dimension; The sparsity constraint is: to encode the latent vector using an autoencoder MEQAE. Apply Norm regularization acts as a penalty, forcing the latent encoding vector The activation values ​​of dimensions above the first set ratio are suppressed to zero or close to zero, and only the features of dimensions below the second set ratio are allowed to remain active; The expression for sparse constraints is as follows: in, Represents the loss function; express The number of divergence terms; Represents the KL divergence term; The weight hyperparameters representing the KL divergence regularization term; Represents prior probability; This represents the probability estimate output by the feature extraction module; express The weight hyperparameters of the sparse constraint term are used to control the latent encoding vector. The strength of sparsity, weight hyperparameter The larger the value, the stronger the constraint on sparsity; KL represents the KL divergence loss. Represents the latent encoding vector of Norm; S332. The purified feature representation is input into the decoder for internal reconstruction calculation to obtain the reconstruction calculation result. At the same time, the purified feature representation is also used as the refined feature representation output by the feature purification module and input into the feature fusion module. The specific expression of the decoder is as follows: in, This indicates that the feature purification module is at time step t Predicted hidden states; This indicates that the feature purification module is at time step t Input; This indicates that the feature purification module is at time step t The final predicted output.

7. The method for diagnosing motor faults in engineering vehicles based on the TAEB model according to claim 6, characterized in that, S34 specifically includes the following steps: S341. After receiving the refined feature representation and temporal features, the feature fusion module uses the bidirectional long short-term memory network inside the feature fusion module to perform temporal modeling on the refined features in order to capture the sequential dependencies of the refined features. At the same time, it enhances the internal association of the global features through the self-attention mechanism to obtain the enhanced global features. The expression for a bidirectional long short-term memory network is as follows: in, This indicates that the bidirectional long short-term memory network is in time step t cellular state; This indicates the forgetting gate in bidirectional long short-term memory networks at time steps. t The output; Indicates time step t Candidate cell state; This indicates the input gate in a bidirectional long short-term memory network at time step. t The output; This indicates element-wise multiplication; This represents the output of the output gate in a bidirectional long short-term memory network; S342. The enhanced global features and temporal features are concatenated and dynamically weighted through a learnable attention fusion layer. Adaptive weighted fusion of local dynamics and global context is then performed to obtain the fused features. S343. The fused features are fed into a fully connected layer for nonlinear transformation and Dropout regularization to output a comprehensive feature vector.

8. The method for diagnosing motor faults in engineering vehicles based on the TAEB model according to claim 7, characterized in that, The total loss function in S43 is as follows: in, Represents the total loss function; Represents cross-entropy loss; This represents the loss from knowledge distillation. This represents the regularization coefficient, used to balance the weights of cross-entropy loss and knowledge distillation loss; In the formula, knowledge distillation loss The calculation formula is as follows: in, Indicates the target distribution; Indicates the predicted distribution; and The teacher model and the student model are respectively in the 1st month. Output probabilities for each category, This indicates all fault categories. Perform a summation calculation.

9. A fault diagnosis system for an engineering vehicle motor, characterized in that, The method for diagnosing motor faults in engineering vehicles as described in any one of claims 1 to 8 is configured or executed.

Images