AGV fault prediction method based on Sophia optimizer and CNN-LSTM model

Abstract

The invention relates to an AGV fault prediction method based on a Sophia optimizer and a CNN-LSTM model. The AGV fault prediction method comprises the steps of obtaining an original audio signal of an AGV in an operation process; performing empirical mode decomposition on the original audio signal to obtain a plurality of IMF components; performing data reconstruction on the plurality of IMF components by using the Pearls correlation coefficient and the variance contribution rate, and screening out an optimal IMF component; establishing a CNN-LSTM model, and training the CNN-LSTM model by using a Sophia optimizer to obtain a fault prediction model; and inputting the optimal IMF component into a fault prediction model, and performing fault prediction on the AGV by using the fault prediction model to obtain a fault prediction result. According to the embodiment of the invention, the original audio signal is analyzed and processed to obtain the optimal IMF component capable of representing the real characteristics of the original audio signal of the AGV, and then the fault prediction model trained by the Sophia optimizer is utilized to carry out fault prediction on the AGV, so that the accuracy and real-time performance of fault detection are improved.

Description

Technical Field

[0001] The present application relates to the technical field of fault prediction, and in particular to an AGV fault prediction method based on a Sophia optimizer and a CNN-LSTM model. Background Art

[0002] An automated guided vehicle (AGV) is a transport vehicle equipped with an electromagnetic or optical automatic guidance device that can travel along a specified guidance path and has safety protection and various transfer functions.

[0003] In modern intelligent manufacturing systems, AGVs, as core equipment for material handling, are widely used in various production workshops. The normal operation of AGVs can ensure the efficient operation of the workshop, but various mechanical failures may occur during the operation of AGVs. If these mechanical failures are not discovered in time, they may cause the production line in the workshop to stagnate or even cause safety accidents. At present, fault detection for AGVs is generally achieved through manual regular troubleshooting, but this method often has problems such as untimely detection and inaccurate fault detection, which affects the efficient operation of the workshop.

[0004] Therefore, it is necessary to provide a new technical solution to improve one or more problems existing in the above solutions.

[0005] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present application, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0006] The purpose of the embodiments of the present application is to provide an AGV fault prediction method based on the Sophia optimizer and the CNN-LSTM model to improve the accuracy and real-time performance of fault detection.

[0007] According to an embodiment of the present application, an AGV fault prediction method based on a Sophia optimizer and a CNN-LSTM model is provided, including: Obtain the original audio signal of AGV during operation; Performing empirical mode decomposition on the original audio signal to obtain multiple IMF components; Reconstructing data of the multiple IMF components by using the Pierce correlation coefficient and variance contribution rate, and selecting the preferred IMF components; A CNN-LSTM model is established, and the CNN-LSTM model is trained using the Sophia optimizer to obtain a fault prediction model; The preferred IMF component is input into the fault prediction model, and the fault prediction model is used to predict the fault of the AGV to obtain a fault prediction result.

[0008] In an exemplary embodiment of the present application, the data of the IMF components are reconstructed by using the Pierce correlation coefficient and the variance contribution rate, and the preferred IMF components are screened out, including: The correlation coefficient between the IMF component and the original audio signal is calculated using formula (1): (1) in, represents the original audio signal collected at the signal sampling point n, express The arithmetic mean of Indicates the sequence number of the IMF component, Indicates IMF components, express The arithmetic mean of Indicates The correlation coefficient between the IMF components and the original audio signal, N Represents the total number of signal sampling points, Indicates the serial number of the signal sampling point; The contribution of the IMF component to the original audio signal is calculated using formula (2): (2) in, Indicates The contribution of each IMF component to the original audio signal. Indicates The variance of the IMF components, represents the sum of variances of all IMF components, Represents the total number of IMF components obtained after the original audio signal is subjected to empirical mode decomposition; The IMF components are screened according to the correlation coefficient and the contribution degree to obtain the preferred IMF components.

[0009] In an exemplary embodiment of the present application, the screening process of the IMF components according to the correlation coefficient and the contribution degree to obtain the preferred IMF components includes: The first indicator threshold and the second indicator threshold are calculated according to formula (3) and formula (4) respectively: (3) (4) in, is the first indicator threshold, is the maximum correlation coefficient, is the second indicator threshold, For the maximum contribution; If the correlation coefficient is greater than the first indicator threshold and the contribution degree is greater than the second indicator threshold, the IMF component is regarded as the preferred IMF component.

[0010] In an exemplary embodiment of the present application, the CNN-LSTM model includes an input layer, a CNN layer, an LSTM layer, an attention mechanism layer, a fully connected layer and an output layer; The input layer is used to receive the preferred IMF component; The CNN layer is used to perform convolution and pooling processing on the preferred IMF component to extract a local feature vector; The LSTM layer is used to capture the long-term dependency of the original audio signal based on the local feature vector and generate a unit output; The attention mechanism layer is used to perform attention weighting processing on the unit output to obtain a weighted feature vector; The fully connected layer is used to perform a linear transformation on the weighted feature vector to obtain a linearly transformed feature vector; The output layer is used to perform data processing on the linearly transformed feature vector to generate a fault prediction result.

[0011] In an exemplary embodiment of the present application, the training of the CNN-LSTM model using the Sophia optimizer includes: Selecting some of the preferred IMF components as training samples, and inputting the training samples into the CNN-LSTM model; The CNN-LSTM model is trained. During the training iteration, the model loss function is calculated using formula (5): (5) in, represents the model loss function, are the model parameters to be updated, Output results for the model. is the dataset label, represents the number of training samples, represents the sequence number of the training sample, , For the The dataset labels corresponding to the training samples are For the 'Model output results obtained after training with ' training samples; Based on the model loss function, the model parameters of the CNN-LSTM model are updated using the Sophia optimizer: (6) (7) (8) (9) in, represents the number of training iterations, Indicates the number of training iterations is When the model parameters, Indicates the number of training iterations is The exponential moving average of the gradient at , Indicates the number of training iterations is The estimated value of the diagonal elements of the Hessian matrix when , Represents a constant, and represents the exponential decay rate, represents the learning rate, Indicates the number of training iterations is The gradient of represents partial derivatives, Indicates the number of training iterations is The value of the model loss function when Indicates the number of training iterations is The model parameters when Indicates the number of training iterations is The exponential moving average of the gradient at , Indicates the number of training iterations is The estimated value of the diagonal elements of the Hessian matrix when ; When the value of the model loss function no longer decreases or the number of iterations reaches a set iteration threshold, the training of the CNN-LSTM model is stopped.

[0012] In an exemplary embodiment of the present application, performing empirical mode decomposition on the original audio signal to obtain multiple IMF components includes: Step a, performing computational processing on the original audio signal to obtain a remaining signal; Step b: performing judgment on the remaining signals: If the remaining signal satisfies the IMF component condition, the remaining signal is the IMF component; if the remaining signal does not satisfy the IMF component condition, the remaining signal replaces the original audio signal, and steps a and b are repeated until the IMF component is obtained; Step c: removing the obtained IMF component from the original audio signal to obtain a signal residual component, and performing a judgment on the signal residual component: If the signal residual component is a monotonic function, the empirical mode decomposition is terminated; if the signal residual component is not a monotonic function, the signal residual component replaces the original audio signal, and steps a to c are repeated until the signal residual component is a monotonic function.

[0013] In an exemplary embodiment of the present application, performing computational processing on the original audio signal to obtain a remaining signal comprises the following steps: Performing cubic spline interpolation processing on the original audio signal to obtain an upper envelope signal and a lower envelope signal; Calculating an average envelope signal based on the upper envelope signal and the lower envelope signal; The original audio signal and the average envelope signal are subjected to a difference process to obtain a residual signal.

[0014] In an exemplary embodiment of the present application, the step of performing judgment on the remaining signal includes: The screening threshold value of the remaining signal is calculated according to formula (10): (10) in, Indicates the screening threshold value, N is the total number of signal sampling points, Indicates the sequence number of the signal sampling point, k represents the number of sieving iterations, Indicates the signal sampling point The first k The remaining signal after sieving iterations is Indicates the signal sampling point The first k -The remaining signal after 1 sieving iteration; The screening threshold value is compared with the threshold value, and if the screening threshold value is less than the threshold value, the remaining signal is considered to meet the IMF component condition.

[0015] In an exemplary embodiment of the present application, the original audio signal is subjected to empirical mode decomposition to obtain a plurality of IMF components; The original audio signal after empirical mode decomposition can be expressed as: (11) in, Indicates the sequence number of the signal sampling point, is the original audio signal collected at the signal sampling point n, Indicates the sequence number of the IMF component, For the IMF components For the The residual signal component, Indicates the total number of IMF components obtained after the original audio signal is subjected to empirical mode decomposition.

[0016] The technical solution provided by this application may have the following beneficial effects: In the embodiments of the present application, on the one hand, multiple IMF components are obtained by performing empirical mode decomposition processing on the original audio signal, and then the Pierce correlation coefficient and variance contribution rate are used to reconstruct the data of the multiple IMF components, thereby analyzing and denoising the original audio signal, eliminating invalid components in the IMF components, and obtaining preferred IMF components, which is conducive to characterizing the true characteristics of the original audio signal of the AGV; on the other hand, a CNN-LSTM model is established and the CNN-LSTM model is trained using the Sophia optimizer to obtain a fault prediction model with good generalization ability and robustness, and then the preferred IMF components are input into the fault prediction model, and the fault prediction model is used to predict the fault of the AGV, thereby improving the accuracy and real-time performance of fault detection.

[0017] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present application, and together with the specification are used to explain the principles of the present application. Obviously, the drawings described below are only some embodiments of the present application, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 A flowchart showing the steps of the AGV fault prediction method based on the Sophia optimizer and the CNN-LSTM model in an exemplary embodiment of the present application; Figure 2 The empirical mode decomposition flow chart in the exemplary embodiment of the present application is shown; Figure 3 A schematic diagram showing the basic structure of the LSTM layer in an exemplary embodiment of the present application is shown; Figure 4 A training flow chart of a CNN-LSTM model in an exemplary embodiment of the present application is shown; Figure 5A flowchart of AGV failure prediction in an exemplary embodiment of the present application is shown. DETAILED DESCRIPTION

[0020] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be more comprehensive and complete and fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0021] In addition, the accompanying drawings are only schematic illustrations of the present application and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and thus their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor devices and / or microcontroller devices.

[0022] This example implementation first provides an AGV fault prediction method based on the Sophia optimizer and the CNN-LSTM model. Figure 1 As shown in , the method may include the following steps: Step S101: obtaining the original audio signal of the AGV during operation; Step S102: performing empirical mode decomposition on the original audio signal to obtain multiple IMF components; Step S103: reconstructing data of multiple IMF components using the Pierce correlation coefficient and variance contribution rate, and selecting the preferred IMF components; Step S104: Establish a CNN-LSTM model, and use the Sophia optimizer to train the CNN-LSTM model to obtain a fault prediction model; Step S105: input the optimal IMF component into the fault prediction model, and use the fault prediction model to predict the fault of the AGV to obtain a fault prediction result.

[0023] In the embodiments of the present application, on the one hand, multiple IMF components are obtained by performing empirical mode decomposition processing on the original audio signal, and then the Pierce correlation coefficient and variance contribution rate are used to reconstruct the data of the multiple IMF components, thereby analyzing and denoising the original audio signal, eliminating invalid components in the IMF components, and obtaining preferred IMF components, which is conducive to characterizing the true characteristics of the original audio signal of the AGV; on the other hand, a CNN-LSTM model is established and the CNN-LSTM model is trained using the Sophia optimizer to obtain a fault prediction model with good generalization ability and robustness, and then the preferred IMF components are input into the fault prediction model, and the fault prediction model is used to predict the fault of the AGV, thereby improving the accuracy and real-time performance of fault detection.

[0024] Next, we will refer to Figures 1 to 5 Each step of the above method in this example implementation is described in more detail.

[0025] Before discussing this method, the following abbreviations are explained.

[0026] Empirical Mode Decomposition (EMD) is a new adaptive signal time-frequency processing method, which is particularly suitable for the analysis and processing of nonlinear and non-stationary signals. The intrinsic mode function (IMF) is a function with specific properties in physics. The instantaneous frequency of any point of the intrinsic mode function is meaningful. Any signal is composed of several intrinsic mode functions. If the intrinsic mode functions overlap with each other, a composite signal is formed. The purpose of empirical mode decomposition is to obtain the intrinsic mode function.

[0027] The Pearson correlation coefficient is a statistic used to measure the strength of the linear relationship between two variables. By calculating the Pearson correlation coefficient, the linear relationship between two variables can be determined. The variance contribution rate is a measure of the contribution of a variable to the overall variance. It reflects the importance or influence of the variable in the data set. Among them, variables with high variance contribution rates are considered to be more important because they can better explain the changes in the data set.

[0028] Convolutional Neural Network (CNN) is a feedforward neural network whose artificial neurons can respond to surrounding units within a certain coverage area, and performs well in large image processing. Long short term memory (LSTM) is a special type of recurrent neural network (RNN) that can solve the problem of gradient disappearance and learn long-term dependencies. The attention mechanism can dynamically allocate attention weights according to the importance of different positions in the input sequence, effectively extracting important features.

[0029] In the prior art, the sigmoid function is often used as an activation function of a neural network due to its monotonic increasing and inverse monotonic increasing properties. It maps variables to between 0 and 1, and works better when the feature differences are complex or not very large. The tanh function is a hyperbolic tangent function, a type of hyperbolic function. The Softmax function can also be called a normalized exponential function, which maps the outputs of multiple neurons to the interval (0, 1). The Sophia optimizer uses an efficient random second-order optimization algorithm that aims to significantly reduce the cost of language model pre-training through its simple plug-and-play features.

[0030] In step S101, the original audio signal of the AGV during operation is obtained, which may specifically include obtaining multiple original audio signals of the AGV in normal operation state, motor bearing fault operation state, motor winding fault operation state, motor drive fault operation state and motor stall fault operation state.

[0031] It should be noted that, during data processing, the audio signal data groups corresponding to the various operating states obtained above need to be processed according to the acquisition time to obtain audio signal data groups with time series characteristics; and each audio signal data group needs to be labeled with a corresponding data set label according to each operating state.

[0032] Specifically, in step S101, the audio information of the AGV during operation can be collected by a microphone to obtain the original audio signal of the AGV during operation.

[0033] In step S102, refer to Figure 2 As shown in , after the original audio signal is obtained, the original audio signal can be subjected to empirical mode decomposition, and then decomposed into multiple IMF components step by step, so as to obtain multiple IMF components.

[0034] In one embodiment, performing empirical mode decomposition on the original audio signal to obtain multiple IMF components includes the following steps: Step a, performing computational processing on the original audio signal to obtain a remaining signal; Step b: judging the remaining signal, that is, judging whether the remaining signal meets the IMF component condition: If the remaining signal meets the IMF component condition, the remaining signal is the IMF component; if the remaining signal does not meet the IMF component condition, the remaining signal replaces the original audio signal, and steps a and b are repeated until the IMF component is obtained; Step c: removing the obtained IMF component from the original audio signal to obtain a signal residual component, and performing a judgment on the signal residual component, that is, judging whether the signal residual component is a monotonic function: If the signal residual component is a monotonic function, the empirical mode decomposition is terminated. If the signal residual component is not a monotonic function, the signal residual component replaces the original audio signal, and steps a to c are repeated until the signal residual component is a monotonic function.

[0035] Through the above-mentioned empirical mode decomposition steps, the original audio signal is decomposed step by step to obtain multiple IMF components. This adaptive signal time-frequency processing method can decompose the signal according to the time scale characteristics of the original audio data itself, and is particularly suitable for the analysis and processing of non-stationary audio signals under various fault operating conditions. The main function is to extract the intrinsic vibration mode in the original audio signal so that the original audio signal can better display its time-frequency characteristics. Each IMF component obtained by the above decomposition represents an intrinsic vibration mode in the original audio signal, and has local time-frequency characteristics, which facilitates accurate and comprehensive analysis of the original audio signal of AGV under different operating conditions.

[0036] Specifically, in step a, the original audio signal is subjected to computational processing to obtain the remaining signal, including the following steps: Step a1: Convert the original audio signal Perform cubic spline interpolation processing to obtain the original audio signal The upper envelope signal and the lower envelope signal ; Step a2: Based on the upper envelope signal and the lower envelope signal The average envelope signal is calculated. Specifically, the average envelope signal can be calculated by formula (12): (12) in, is the average envelope signal, is the upper envelope signal, is the lower envelope signal; Step a3: The original audio signal can be converted into Subtract the average envelope signal to get the remaining signal: (13) in, For the remaining signals, At the signal sampling point The original audio signal collected at is the average envelope signal.

[0037] The above calculation steps illustrate how to calculate and process the original audio signal to obtain the residual signal, and the above steps are performed at least once in each empirical mode decomposition process. It should be noted that the original audio signal, the residual signal, and the average envelope signal are not subscripted, and the above calculation formula is applicable to any empirical mode decomposition process.

[0038] Specifically, the remaining signals are judged in step b, including: The screening threshold of the remaining signal is calculated according to formula (10): (10) in, Indicates the screening threshold value, N is the total number of signal sampling points, Indicates the sequence number of the signal sampling point, k represents the number of sieving iterations, Indicates the signal sampling point The first k The remaining signal after sieving iterations is Indicates the signal sampling point The first k -The remaining signal after 1 sieving iteration; The screening threshold value is compared with the threshold value. If the screening threshold value is less than the threshold value, the remaining signal is considered to meet the IMF component condition.

[0039] The threshold value can be set to 0.25. The screening threshold value calculated by formula (10) can be compared with the threshold value set to 0.25 to determine whether the remaining signal meets the IMF component condition.

[0040] It should be noted that repeating step a to step c once can be understood as one empirical mode decomposition, and by repeating step a to step c multiple times, multiple decompositions of the original audio signal can be achieved.

[0041] It should also be noted that the original audio signal generally needs to undergo multiple empirical mode decompositions to obtain multiple IMF components before the remaining signal residual component is a monotonic function. When the signal residual component is a monotonic function, the signal residual component becomes smooth and there is no decomposable oscillation component in the signal residual component, so the empirical mode decomposition is no longer continued.

[0042] In this embodiment, by repeating steps a to c multiple times, empirical mode decomposition is performed on the original audio signal to obtain multiple IMF components; Among them, the original audio signal after empirical mode decomposition can be expressed as: (11) in, Indicates the sequence number of the signal sampling point, is the original audio signal collected at the signal sampling point n, Indicates the sequence number of the IMF component, For the IMF components For the The residual signal component, Indicates the total number of IMF components obtained after the original audio signal is subjected to empirical mode decomposition.

[0043] Above Can be regarded as i The first empirical mode decomposition It should be noted that for the multiple IMF components obtained through empirical mode decomposition, although the intrinsic vibration modes in the original audio signal are extracted, so that the original audio signal can better show its time-frequency characteristics, there are two parts in these multiple IMF components: effective components and invalid components. In order to characterize the true original audio signal characteristics of AGV, it is necessary to fully eliminate the invalid components.

[0044] Therefore, the present application further introduces the Pearson correlation coefficient and variance contribution rate through step S103 to optimize the selection of multiple IMF components.

[0045] In step S103, the IMF components are reconstructed using the Pierce correlation coefficient and the variance contribution rate, and the preferred IMF components are screened out, which may include the following steps: Step S1031: Calculate the correlation coefficient between the IMF component and the original audio signal using formula (1): (1) in, represents the original audio signal collected at the signal sampling point n, express The arithmetic mean of Indicates the sequence number of the IMF component, Indicates IMF components, express The arithmetic mean of Indicates The correlation coefficient between the IMF components and the original audio signal, N Represents the total number of signal sampling points, Indicates the serial number of the signal sampling point; Step S1032: Calculate the contribution of the IMF component to the original audio signal using formula (2): (2) in, Indicates The contribution of each IMF component to the original audio signal. Indicates The variance of the IMF components, represents the sum of variances of all IMF components, Represents the total number of IMF components obtained after the original audio signal is subjected to empirical mode decomposition; Step S1033: Screening the IMF components according to the correlation coefficient and contribution degree to obtain the preferred IMF components.

[0046] Through the above steps, invalid components in multiple IMF components can be eliminated, so that the obtained preferred IMF components only retain valid components, and the denoising process of the original audio signal is realized, thereby characterizing the real original audio signal characteristics of the AGV.

[0047] Furthermore, the IMF components are screened according to the correlation coefficient and contribution degree to obtain the preferred IMF components, including: The first indicator threshold and the second indicator threshold are calculated according to formula (3) and formula (4) respectively: (3) (4) in, is the first indicator threshold, is the maximum correlation coefficient, is the second indicator threshold, For the maximum contribution; If the correlation coefficient calculated according to formula (1) is greater than the first indicator threshold, and the contribution degree calculated according to formula (2) is greater than the second indicator threshold, the IMF component is regarded as the preferred IMF component.

[0048] It should be noted that the above is the maximum correlation coefficient between the IMF component and the original audio signal calculated using formula (1), The maximum contribution degree among the contributions of the IMF component to the original audio signal is calculated using formula (2).

[0049] In step S104, a CNN-LSTM model is established, which includes an input layer, a CNN layer, an LSTM layer, an attention mechanism layer, a fully connected layer and an output layer; The input layer is used to receive the preferred IMF components; The CNN layer is used to perform convolution and pooling on the preferred IMF components to extract local feature vectors; The LSTM layer is used to capture the long-term dependencies of the original audio signal based on the local feature vector and generate unit outputs; The attention mechanism layer is used to perform attention weighting on the unit output to obtain a weighted feature vector; The fully connected layer is used to perform linear transformation on the weighted feature vector to obtain the linearly transformed feature vector; The output layer is used to perform data processing on the feature vector after linear transformation to generate fault prediction results.

[0050] The CNN-LSTM model of the present application combines the respective characteristics of the CNN layer and the LSTM layer, which can not only make full use of the time-frequency characteristics of the original audio signal of the AGV, but also extract the local features of the original audio signal through the CNN layer, and capture the long-term dependencies of the original audio signal through the timing modeling function of the LSTM layer to improve the accuracy of classification and prediction of fault types; in addition, an attention mechanism layer is added after the LSTM layer, which can further process the output of the LSTM layer. Through the attention mechanism, the CNN-LSTM model can dynamically focus on the signal features that are most representative of faults at different time steps, and then lock the moment when the fault exists, thereby improving the accuracy of fault prediction.

[0051] It is important to understand that, assuming that the original audio signal of the device will have specific frequency features directly related to the fault state (such as increased or decreased frequency) before the fault occurs. This specific frequency feature can reflect the precursor of the device failure. Although this specific frequency may not be obvious in a short period of time, it will gradually accumulate over time, thus reflecting the potential abnormal trend in the operation of the device. By capturing these specific frequency features directly related to the fault state and their long-term dependencies, the LSTM layer can identify fault signs in advance before the fault occurs, thereby achieving accurate fault prediction.

[0052] In the above-mentioned fully connected layer, each neuron of the fully connected layer is connected to all outputs of the attention mechanism layer to form a densely connected network structure. Through this densely connected network structure, the fully connected layer can perform a linear transformation on the weighted feature vector and convert the weighted feature vector into the format required by the output layer to generate fault prediction results.

[0053] In one embodiment, the CNN-LSTM model includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, an LSTM layer, an attention mechanism layer, a fully connected layer and an output layer connected in sequence. Among them, the first convolutional layer, the first pooling layer, the second convolutional layer and the second pooling layer belong to the CNN layer. The IMF component can be input into the CNN-LSTM model through the input layer; the first convolutional layer is used to preliminarily extract local features in the IMF component; the first pooling layer is used to reduce the size of the feature map and retain the most important features in the IMF component; the input of the second convolutional layer is the feature map after pooling by the first pooling layer, so the second convolutional layer can further extract higher-level features in the audio signal; the second pooling layer is used to further compress the feature map and reduce the dimension of the audio data, thereby ensuring that the subsequent LSTM layer can efficiently process these local features, and it is also conducive to reducing the computational complexity; the LSTM layer, the attention mechanism layer, the fully connected layer and the output layer have the same specific functions as those in the aforementioned step S104.

[0054] Specifically, refer to Figure 3 As shown in , the LSTM layer in this embodiment can be composed of four parts, namely, input gate, forget gate, output gate and cell state, where the input gate determines how much new information is added to the cell state, the forget gate determines what information is discarded from the cell state, and the output gate determines what information is finally output. The cell state defines and maintains an internal memory unit state in the entire cycle. The update formulas for the input gate, forget gate, output gate and cell state at each time step in the neuron cells of the LSTM layer are as follows: Forget Gate: (14) Input Gate: (15) Candidate status for input: (16) Output of memory cell: (17) Output Gate: (18) Unit output: (19) in, represents the time step of the LSTM layer input vector, 、 、 、 Represent the weight matrices of the input gate, forget gate, output gate, and candidate states of the input respectively; 、 、 、 Respectively represent the bias items of the input gate, forget gate, output gate and input candidate states; express The unit output at the time step, express The output of the memory cell at time step; express The input vector at the time step; represents the hyperbolic tangent activation function; represents a nonlinear activation function; , , , , , Respectively Forget gate at time step, input gate, candidate state of input, output of memory cell, output gate, unit output.

[0055] It should be noted that Figure 3 In Represents the sigmoid nonlinear activation function.

[0056] Furthermore, after the output of the LSTM layer, the attention mechanism layer is used to perform weighted summation on the unit output of the LSTM layer, and finally the weighted summed unit output is passed to the fully connected layer. The calculation of the attention mechanism layer is divided into two steps. The first step is to calculate the weight based on the unit output. The calculation formula is: (20) The second step is to use the weights calculated in the first step to perform weighted summation on the unit outputs. The calculation formula is: (twenty one) in, represents the weight, represents the unit output, represents the target vector, represents the score function, represents an exponential function with the natural constant e as the base, represents the weighted sum of unit outputs, Represents the total number of time steps.

[0057] In step S104, refer to Figure 4 As shown in , using the Sophia optimizer to train the CNN-LSTM model may include the following process: Select some preferred IMF components as training samples, such as 80% of the preferred IMF components as training samples, and input the training samples into the CNN-LSTM model; When training the CNN-LSTM model, during the training iteration, the root mean square error between the model output and the data group label is used as the model loss function, and the model loss function is calculated using formula (5): (5) in, represents the model loss function, are the model parameters to be updated, Output results for the model. is the dataset label, represents the number of training samples, represents the sequence number of the training sample, , For the The dataset labels corresponding to the training samples are For the The model output results obtained after training with training samples; Based on the model loss function, the Sophia optimizer is used to update the model parameters of the CNN-LSTM model: (6) (7) (8) (9) in, represents the number of training iterations, Indicates the number of training iterations is When the model parameters, Indicates the number of training iterations is The exponential moving average of the gradient at , Indicates the number of training iterations is The estimated value of the diagonal elements of the Hessian matrix when , Represents a constant, and represents the exponential decay rate, represents the learning rate, Indicates the number of training iterations is The gradient of represents partial derivatives, Indicates the number of training iterations is The value of the model loss function when Indicates the number of training iterations is The model parameters when Indicates the number of training iterations is The exponential moving average of the gradient at , Indicates the number of training iterations is The estimated value of the diagonal elements of the Hessian matrix when ; When the value of the model loss function no longer decreases or the number of iterations reaches the set iteration threshold, stop training the CNN-LSTM model; Stores the model parameters of the CNN-LSTM model.

[0058] When the value of the model loss function continues to decrease and the number of iterations does not reach the set iteration threshold, continue to train the CNN-LSTM model.

[0059] This application uses the Sophia optimizer to train the CNN-LSTM model, which can adjust the learning rate more accurately, thereby accelerating the convergence of the model. The Sophia optimizer introduces a clipping mechanism to ensure that the update step size remains stable on complex non-convex loss functions.

[0060] It should be noted that before training the CNN-LSTM model, it is also necessary to set the initial model parameters of the CNN-LSTM model, the initial learning rate of the iterative training process and other parameters.

[0061] The above exponential decay rate The value can be 0.9, exponential decay rate The value can be 0.999, the initial learning rate The acceptable value is 0.001.

[0062] In step S105, the fault prediction model is used to predict the fault of the AGV to obtain a fault prediction result. Figure 5As shown in , when using the fault prediction model to predict the fault of AGV, the original audio signal is first obtained, and the original audio signal is subjected to empirical mode decomposition to obtain multiple IMF components. The collected original audio signal is analyzed and denoised using the Pierce correlation coefficient and variance contribution rate to remove invalid components and obtain the preferred IMF components. The preferred IMF components are then input into the fault prediction model, and the preferred IMF components are processed through the input layer, CNN layer, LSTM layer, attention mechanism layer, fully connected layer and output layer in the fault prediction model to achieve fault prediction and obtain the fault prediction result.

[0063] Specifically, the output layer of the fault prediction model includes a Dropout layer and a Softmax layer. The Dropout layer is used to randomly discard a certain proportion of neurons during the training process to reduce the model's dependence on specific neurons, thereby improving the generalization and robustness of the model. Through this mechanism, the model can better adapt to different input data and reduce the risk of overfitting. The Softmax layer is used to convert the output of the neural network into a probability distribution to ensure the normalization of the output results. Specifically, the Softmax layer converts the output value of each output unit into a probability form through the Softmax function. The calculation formula of the Softmax function is: (twenty two) in, represents the fault prediction result obtained by using the fault prediction model. Represents the neural network j The output value of the output unit, , Represents the output value of each output unit of the neural network. The output format of the Softmax layer is = [ ],in, represents the probability of the corresponding AGV normal operation state, represents the probability of the corresponding motor bearing fault operating state, represents the probability of the corresponding motor winding fault operating state, represents the probability of the corresponding motor drive fault operating state, Indicates the probability of the corresponding motor stall fault running state. Among them, the running state and probability value of the AGV corresponding to the maximum value in the output result of the Softmax layer is the fault prediction result.

[0064] In summary, the AGV fault prediction method based on Sophia optimizer and CNN-LSTM model proposed in this application, combined with the respective characteristics of empirical mode decomposition, CNN layer and LSTM layer, can make full use of the time-frequency characteristics of the original audio signal of AGV, by performing empirical mode decomposition on the original audio signal, extracting the local features of the original audio signal through the CNN layer, and capturing the long-term dependencies of the original audio signal through the time series modeling of the LSTM layer, to achieve accurate prediction of AGV operation failures, and introduce an attention mechanism layer so that the CNN-LSTM model can selectively focus on the most important part of the original audio signal, further improving the CNN-LSTM model's ability to distinguish different fault types. This mechanism not only helps to improve the accuracy of the CNN-LSTM model, but also enhances the generalization ability of the CNN-LSTM model, reduces the risk of overfitting, and makes fault detection more accurate and efficient. In addition, this application also uses the Sophia optimizer to train the CNN-LSTM model. The Sophia optimizer uses the diagonal approximation information of the Hessian matrix to more accurately adjust the learning rate, thereby accelerating the convergence of the model. The Sophia optimizer introduces a clipping mechanism to ensure that the update step size remains stable on complex non-convex loss functions. This mechanism enables the Sophia optimizer to effectively avoid converging to local optimal or oscillating problems when facing rapidly changing or highly non-convex loss surfaces.

[0065] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary technical means in the art that are not disclosed in the present application.

Claims

1. An AGV fault prediction method based on Sophia optimizer and CNN-LSTM model, characterized in that: include: Obtain the original audio signal of AGV during operation; Performing empirical mode decomposition on the original audio signal to obtain multiple IMF components; Reconstructing data of the multiple IMF components by using the Pierce correlation coefficient and the variance contribution rate, and selecting the preferred IMF components; A CNN-LSTM model is established, and the CNN-LSTM model is trained using the Sophia optimizer to obtain a fault prediction model; The preferred IMF component is input into the fault prediction model, and the fault prediction model is used to predict the fault of the AGV to obtain a fault prediction result.

2. The AGV fault prediction method based on Sophia optimizer and CNN-LSTM model according to claim 1, characterized in that: The method of reconstructing the data of the IMF components by using the Pierce correlation coefficient and the variance contribution rate, and selecting the preferred IMF components, includes: The correlation coefficient between the IMF component and the original audio signal is calculated using formula (1): （1） in, represents the original audio signal collected at the signal sampling point n, express The arithmetic mean of Indicates the sequence number of the IMF component, Indicates IMF components, express The arithmetic mean of Indicates The correlation coefficient between the IMF components and the original audio signal, N Represents the total number of signal sampling points, Indicates the serial number of the signal sampling point; The contribution of the IMF component to the original audio signal is calculated using formula (2): （2） in, Indicates The contribution of each IMF component to the original audio signal. Indicates The variance of the IMF components, represents the sum of variances of all IMF components, Represents the total number of IMF components obtained after the original audio signal is subjected to empirical mode decomposition; The IMF components are screened according to the correlation coefficient and the contribution degree to obtain the preferred IMF components.

3. The AGV fault prediction method based on Sophia optimizer and CNN-LSTM model according to claim 2 is characterized in that: The screening process of the IMF components according to the correlation coefficient and the contribution degree to obtain the preferred IMF components includes: The first indicator threshold and the second indicator threshold are calculated according to formula (3) and formula (4) respectively: （3） （4） in, is the first indicator threshold, is the maximum correlation coefficient, is the second indicator threshold, For the maximum contribution; If the correlation coefficient is greater than the first indicator threshold and the contribution degree is greater than the second indicator threshold, the IMF component is regarded as the preferred IMF component.

4. The AGV fault prediction method based on Sophia optimizer and CNN-LSTM model according to claim 1, characterized in that: The CNN-LSTM model includes an input layer, a CNN layer, an LSTM layer, an attention mechanism layer, a fully connected layer and an output layer; The input layer is used to receive the preferred IMF component; The CNN layer is used to perform convolution and pooling processing on the preferred IMF component to extract a local feature vector; The LSTM layer is used to capture the long-term dependency of the original audio signal based on the local feature vector and generate a unit output; The attention mechanism layer is used to perform attention weighting processing on the unit output to obtain a weighted feature vector; The fully connected layer is used to perform a linear transformation on the weighted feature vector to obtain a linearly transformed feature vector; The output layer is used to perform data processing on the linearly transformed feature vector to generate a fault prediction result.

5. The AGV fault prediction method based on Sophia optimizer and CNN-LSTM model according to claim 1, characterized in that: The method of training the CNN-LSTM model using the Sophia optimizer includes: Selecting some of the preferred IMF components as training samples, and inputting the training samples into the CNN-LSTM model; The CNN-LSTM model is trained. During the training iteration, the model loss function is calculated using formula (5): （5） in, represents the model loss function, are the model parameters to be updated, Output results for the model. is the dataset label, represents the number of training samples, represents the sequence number of the training sample, , For the The dataset labels corresponding to the training samples are For the 'Model output results obtained after training with ' training samples; Based on the model loss function, the model parameters of the CNN-LSTM model are updated using the Sophia optimizer: （6） （7） （8） （9） in, represents the number of training iterations, Indicates the number of training iterations is When the model parameters, Indicates the number of training iterations is The exponential moving average of the gradient at , Indicates the number of training iterations is The estimated value of the diagonal elements of the Hessian matrix when , Represents a constant, and represents the exponential decay rate, represents the learning rate, Indicates the number of training iterations is The gradient of represents partial derivatives, Indicates the number of training iterations is The value of the model loss function when Indicates the number of training iterations is The model parameters when Indicates the number of training iterations is The exponential moving average of the gradient at , Indicates the number of training iterations is The estimated value of the diagonal elements of the Hessian matrix when ; When the value of the model loss function no longer decreases or the number of iterations reaches a set iteration threshold, the training of the CNN-LSTM model is stopped.

6. The AGV fault prediction method based on Sophia optimizer and CNN-LSTM model according to claim 1, characterized in that: The performing empirical mode decomposition on the original audio signal to obtain multiple IMF components includes: Step a, performing computational processing on the original audio signal to obtain a remaining signal; Step b: performing judgment on the remaining signals: If the remaining signal satisfies the IMF component condition, the remaining signal is the IMF component; if the remaining signal does not satisfy the IMF component condition, the remaining signal replaces the original audio signal, and steps a and b are repeated until the IMF component is obtained; Step c: removing the obtained IMF component from the original audio signal to obtain a signal residual component, and performing a judgment on the signal residual component: If the signal residual component is a monotonic function, the empirical mode decomposition is terminated; if the signal residual component is not a monotonic function, the signal residual component replaces the original audio signal, and steps a to c are repeated until the signal residual component is a monotonic function.

7. The AGV fault prediction method based on Sophia optimizer and CNN-LSTM model according to claim 6, characterized in that: The performing of computation processing on the original audio signal to obtain a remaining signal comprises the following steps: Performing cubic spline interpolation processing on the original audio signal to obtain an upper envelope signal and a lower envelope signal; Calculating an average envelope signal based on the upper envelope signal and the lower envelope signal; The original audio signal and the average envelope signal are subjected to a difference process to obtain a residual signal.

8. The AGV fault prediction method based on Sophia optimizer and CNN-LSTM model according to claim 6, characterized in that: The performing judgment on the remaining signal comprises: The screening threshold value of the remaining signal is calculated according to formula (10): （10） in, Indicates the screening threshold value, N is the total number of signal sampling points, Indicates the sequence number of the signal sampling point, k represents the number of sieving iterations, Indicates the signal sampling point The first k The remaining signal after sieving iterations is Indicates the signal sampling point The first k -The remaining signal after 1 sieving iteration; The screening threshold value is compared with the threshold value, and if the screening threshold value is less than the threshold value, the remaining signal is considered to meet the IMF component condition.

9. The AGV fault prediction method based on Sophia optimizer and CNN-LSTM model according to any one of claims 1 to 8, characterized in that: The method further comprises performing empirical mode decomposition on the original audio signal to obtain a plurality of IMF components; The original audio signal after empirical mode decomposition can be expressed as: （11） in, Indicates the sequence number of the signal sampling point, is the original audio signal collected at the signal sampling point n, Indicates the sequence number of the IMF component, For the IMF Components For the The residual signal component, Indicates the total number of IMF components obtained after the original audio signal is subjected to empirical mode decomposition.

Images