Elevator fault diagnosis method based on short sequence temporal convolutional network

Abstract

An elevator fault diagnosis method based on a short sequence time convolutional network belongs to the technical field of elevator fault diagnosis and vibration fault diagnosis. First, a one-dimensional convolution method is used to reduce the dimension of the original time domain vibration signal of the elevator; then, a time convolutional neural network is used in combination with a neural ordinary differential equation to realize the extraction of local fault features of a short sequence; finally, a self-attention mechanism is introduced to extract richer global fault-related feature information for elevator fault diagnosis. The present invention provides an elevator fault diagnosis method based on a short sequence time convolutional network that effectively improves the accuracy and reliability of fault detection.

Description

Technical Field

[0001] The invention belongs to the field of elevator fault diagnosis and vibration fault diagnosis, and specifically relates to a short sequence time convolution network elevator fault diagnosis method. Background Art

[0002] With the improvement of computing power and advanced network training technology, deep learning has been widely used in the field of fault diagnosis. As the most commonly used special equipment in people's lives, the operating status of elevators is directly related to the functionality of buildings and the safety of users. The diagnosis of elevator car vibration signals has always been a difficult problem. The temporal convolutional neural network was proposed by Shaojie Bai et al. in 2018. It has now become the mainstream method for processing sequence data, and the temporal convolutional neural network has a good effect in processing fault diagnosis of vibration signals.

[0003] Deep learning shows great advantages in fault diagnosis methods because it is able to learn features from large amounts of data. And by building a neural network with multiple hidden layers through a deep learning architecture, hierarchical representations can be learned from raw data. In machine fault diagnosis, deep architectures can automatically learn useful representations and identify certain fault modes from large amounts of sensor data. For example, INCE et al. used an overlapping segmentation preprocessing method to construct training and test samples of various fault states of motors, and divided each training sample into "time steps" at a certain scale as the input of the established one-dimensional convolutional neural network model. Zheng et al. used the TCN model to extract fault signal features through dilated convolution, dropout and residual structure, and realized the identification of different fault types of rolling bearings; Zhang et al. proposed a fault diagnosis method based on multi-scale temporal convolutional network. According to the multi-scale characteristics and time dependence of industrial process monitoring data, multi-scale time-frequency information was extracted by discrete wavelet transform to represent the original data, and then combined with the time convolutional network to capture long-term temporal features from the sequence processing data; Sun Yu et al. divided the vibration data into multiple sub-blocks, and used TCN to mine the features of the chiller data, and strengthened the influence weight of key features on the fault diagnosis results through the self-attention layer to realize fault diagnosis.

[0004] Neural ODEs are a class of methods that use neural network structures to parameterize solutions to ordinary differential equations. Since Chen et al. proposed them in 2018, they have received widespread attention in dynamic system modeling. Dupont et al. developed an improved Neural ODE combined with a latent variable model to improve the modeling accuracy of dynamic systems; Lu et al. explored a method that combined graph convolutional networks and Neural ODEs for modeling the dynamics of complex network structures; Zhong et al. proposed a framework that integrates adaptive control theory with Neural ODEs, aiming to optimize the control strategy of nonlinear dynamic systems. These studies have shown that Neural ODEs show great potential and adaptability in solving short sequence dynamic modeling problems.

[0005] Although the above methods adjust the neural network structure for specific problems or directly obtain fault diagnosis knowledge in the target domain, and have made certain contributions to the field of deep learning fault diagnosis, the current challenges in the field of elevator fault diagnosis include: (1) Most neural network models only consider long-sequence features. Since the vibration signal of the elevator shows different feature distributions at different time scales, feature extraction that relies solely on long-sequence data is usually unable to fully characterize its health status. (2) When processing complex elevator fault data sets, traditional time convolutional network models often find it difficult to quickly filter out key information from huge data. Summary of the invention

[0006] In order to overcome the shortcomings of the prior art, the present invention proposes an elevator fault diagnosis method based on a short-sequence temporal convolutional network that effectively improves the accuracy and reliability of fault detection. First, a one-dimensional convolution method is used to reduce the dimension of the original time-domain vibration signal of the elevator. Then, a temporal convolutional neural network is combined with a neural ordinary differential equation to realize the extraction of local fault features of short sequences. Finally, a self-attention mechanism is introduced to extract richer global fault-related feature information. On this basis, the present invention constructs an improved temporal convolutional network, uses a self-attention mechanism to integrate local and overall features, and designs a short-sequence temporal convolutional network (Self-Attention Temporal Convolutional Neural Ordinary Differential Equation, SA-TCNODE) ​​for elevator fault diagnosis.

[0007] In order to solve the above technical problems, the present invention proposes an elevator fault diagnosis method based on a short sequence time convolutional network, comprising the following steps:

[0008] The first step is to collect the original vibration signal of the elevator car running acceleration: install an acceleration sensor in the elevator car, and realize remote data acquisition through the TCP / IP communication protocol; use the serial port connection method to connect the acceleration sensor, elevator and gateway, and use Modbus and RS485 protocols for data communication;

[0009] The second step is to obtain the original vibration signal. The processing process is as follows:

[0010] Step (2.1) calculates the time domain characteristic index and adjusts the elevator car vibration acceleration signal by using the time domain analysis method;

[0011] Step (2.2) segments the vibration signal processed in step (2.1), samples it using a sliding window containing M data points, and then slides the window backward by a number of data points greater than M to intercept the original signal, and finally obtains a valid original vibration signal;

[0012] Step (2.3) reduces the dimensionality of the signal through one-dimensional convolution;

[0013] Step (2.4) performs maximum pooling on the signal obtained by one-dimensional convolution;

[0014] Step (2.5) obtaining the final processed vibration signal;

[0015] The third step is to optimize the TCN model, introduce gated convolution and neural ordinary differential equations into dilated convolution and residual connection to form an improved TCNODE model. The process is as follows:

[0016] Step (3.1) The dilated convolution in TCNODE introduces gated convolution:

[0017] Step (3.2) improves the residual connection method into a Neural Ordinary Differential Equation method;

[0018] Step 4: Add the self-attention mechanism to form an improved SA-TCNODE model for elevator fault diagnosis. The process is as follows:

[0019] The local feature output sequence of the sub-block in step (4.1) is converted into Y vectors through adaptive linear mapping, and used to generate the query vector Q, key vector K and value vector V;

[0020] Step (4.2) calculates the attention coefficient based on the Q and K vectors;

[0021] Step (4.3) calculates the weight of each sub-block by weighted summing of the attention coefficient value vector V and concatenates the weighted local features to obtain the global features;

[0022] Step (4.4) converts the global features The output is input into the fully connected layer to obtain the output M, and finally the output M is input into the softmax function for normalization to obtain the elevator fault diagnosis result.

[0023] Furthermore, the fault diagnosis method further comprises the following steps:

[0024] Step 5: Introduce the adjoint sensitivity back propagation method to train the model. The process is as follows:

[0025] Step (5.1) uses the cross entropy loss function to train the model;

[0026] The gradient of the loss with respect to the network parameters in step (5.2) is directly calculated using the adjoint function and the ODE solver.

[0027] Furthermore, the process of step (2.1) is as follows:

[0028] Step (2.1.1) calculates the peak-to-peak value in the vibration signal samples , the formula is as follows,

[0029] ;

[0030] ;

[0031] ;

[0032] in, is the sampling value of the vibration acceleration signal, and max is the sampling value The maximum value, min is the sampling value The minimum value of is the maximum value in the signal samples, is the minimum value in the signal samples;

[0033] The peak-to-peak value is an indicator used to describe the size of the signal variation range within a period. For non-periodic car vibration signals, the peak-to-peak value can be used to measure the amplitude or amplitude variation of the signal, that is, the maximum deviation of the signal.

[0034] Step (2.1.2) calculates the average amplitude of the vibration signal samples RMS , the formula is as follows,

[0035] ;

[0036] ;

[0037] Where N is the number of sampling points, ∑ is the summation operation, which is used in mathematics and statistics to add a series of values. for The absolute value of

[0038] Among them, the average amplitude is used to describe the degree of change of the vibration signal caused by the change of the axis position; the root mean square reflects the average energy intensity and stability of the vibration signal;

[0039] Step (2.1.3) calculates the square root amplitude of the vibration signal samples , the formula is as follows,

[0040] ;

[0041] Among them, the square root amplitude is an indicator used to measure the signal amplitude, which can more accurately reflect the energy distribution and overall strength of the signal;

[0042] Step (2.1.4) adjusts the original vibration signal according to the calculated time domain features , the formula is as follows,

[0043] ;

[0044] in, , They are the target peak-to-peak value, target average amplitude, target RMS value, and target RMS amplitude. is the noise signal used to adjust the vibration signal.

[0045] Furthermore, the step (2.2) obtains a valid original vibration signal ; In step (2.3), the convolution kernel is 16, the convolution kernel width is 12, and the step length is 4.

[0046] ;

[0047] in, represents the convolution operation output of the vibration signal corresponding to the i-th feature map, represents the activation function, represents convolution, is a one-dimensional convolution kernel, T is the number of feature maps;

[0048] Among them, one-dimensional convolution is the most widely used method for data dimensionality reduction in convolutional networks;

[0049] In step (2.4), the output after maximum pooling is,

[0050] ;

[0051] Among them, z represents the width of the maximum pooling, and then the pooling result is subjected to the dropout operation. Dropout can randomly discard some neurons to prevent the model from overfitting;

[0052] Among them, the role of the maximum pooling layer is to reduce the number of features and extract key features;

[0053] In the step (2.5), the final processed vibration signal is obtained.

[0054] .

[0055] In the step (3.1), the temporal convolutional network (TCN) is a network model proposed by Shaojie Bai, and its architecture is mainly achieved by changing the convolutional neural network (CNN). Dilated convolution expands the coverage of the convolution kernel by introducing a dilation factor without increasing the number of parameters or computational complexity; residual connection makes it easier to optimize the deep neural network structure by introducing residual blocks, which greatly facilitates the application of deep network structures;

[0056] Among them, gated convolution is a component of an advanced convolutional neural network, which is mainly responsible for controlling the transmission of information flow. The process is as follows:

[0057] Step (3.1.1) uses convolutional layers and loop structures to construct gated convolutions. In gated convolutions, the output of each convolutional layer passes through a gating mechanism that uses a sigmoid activation function to determine the amount of information transferred in each channel. The width of the convolution kernel in the gated convolution layer is , the gating mechanism generates a gating value between 0 and 1 for each channel to dynamically adjust the flow of information, The number of channels is 2, 3, or 6 to cover all cycles of vibration data, and the number of channels is 64, 128, or 256.

[0058] ;

[0059] in, is the gated convolutional layer, and are the weights and biases of the gated convolutional layer, represents the convolution operation, The convolution kernel size is The network parameters when , the sigmoid function is a smooth curve function widely used to map real numbers to the (0,1) interval, and is often used as an activation function in binary classification problems;

[0060] Step (3.1.2) uses convolutional layers and loop structures to construct dilated convolution. Dilated convolution expands the receptive field by increasing the spacing between each unit in the convolution kernel (dilation rate P) without increasing the number of parameters. The dilation rate P is 1, and the convolution kernel width is Choose 2, 3, 6, and the number of channels is 64, 128, 256.

[0061] ;

[0062] in, is the dilated convolutional layer, and are the weights and biases of the dilated convolutional layer, represents the convolution operation, The convolution kernel size is The network parameters when , the tanh function is an S-type activation function that maps real numbers to the (-1,1) interval, and is often used in neural networks to centralize data;

[0063] Step (3.1.3) obtains the convolutional layer formula of TCNODE,

[0064] ;

[0065] in, represents element-wise product, is the dilated convolution output matrix obtained in step (3.1.2), is the output vector, is the gated convolution output matrix obtained in step (3.1.1), is the output vector.

[0066] The process of step (3.2) is as follows:

[0067] Step (3.2.1) TCN extracts and aggregates temporal information in a non-recursive manner by stacking multiple residual convolution blocks. The local feature L of the sub-block is obtained by the following formula:

[0068] ;

[0069] Among them, is the individually parameterized temporal convolutional layer in step (3.1), Represents a truncation function, taking only The last dimension of elements, The sequence length is The output of the tth layer. The residual output The last dimension of must be truncated to before adding its transformation , because the dilated convolution causes the length of the potential representation to gradually decrease, i.e. , ,in, is the initial state, r, k , l , R are the dilation factor, convolution kernel size, step size and model receptive field respectively. To ensure R > T to achieve lossless encoding of all data, the present invention defines that when r=1, , when r>1, ;

[0070] Step (3.2.2) The idea of ​​the Neural ODE is to treat time as the number of layers. By inserting more layers and taking smaller integration steps, we can directly parameterize and approximate the continuous evolution of the underlying state.

[0071] ;

[0072] in, For function The output, For function Parameters;

[0073] Step (3.2.3) converts the propagation depth in TCNODE to Decomposed into integration time and step length The combination of , the magical ordinary differential equation representation of TCNODE can be realized by increasing the variable integration step size dt,

[0074] ;

[0075] Step (3.2.4) for the termination time and the initial state , through the order , a set of parameters Θ is used to characterize the entire continuous time aggregation process, and the local features L of the sub-blocks are extracted.

[0076] ;

[0077] in, The integration process is used to solve the cumulative amount of a function on a specific interval. It is a black-box ODE solver used to solve neural ordinary differential equations.

[0078] In the step (4.1), the local feature output sequence of the sub-block is Through adaptive linear mapping, it is converted into Y length The vector is used to generate the query vector Q, key vector K and value vector V. The generation formula of Q, K and V is as follows:

[0079] ;

[0080] in, , , It is a learnable weight matrix, which can increase the expressiveness of the model, so that the calculated weight matrix has higher generalization ability;

[0081] In the step (4.2), the attention coefficient is calculated based on the Q and K vectors ,

[0082] ;

[0083] in The probability represents the degree of association between two vectors, e is the base of the natural logarithm, ' is the transposition number, and the softmax function is a normalized exponential function in multi-class classification problems;

[0084] In step (4.3), the global feature G is expressed as,

[0085] ;

[0086] In step (4.4), normalization is expressed as,

[0087] ;

[0088] Among them, c is the dimension of the softmax function input, is the diagnosis object category probability.

[0089] In step (5.1), the cross entropy loss function is used to train the model. The calculation formula is as follows:

[0090] ；

[0091] in, represents the actual label, N represents the total number of samples, is the inverse operation of exponential operation, For the sum operation, the Adam method is finally used as the optimizer to optimize the weights and bias terms;

[0092] In the step (5.2), the loss Network parameters The gradient of is directly calculated by the adjoint function and the ODE solver,

[0093] ;

[0094] in, is the adjoint matrix.

[0095] The beneficial effects of the present invention are mainly manifested in:

[0096] 1. The internal structure of the traditional TCN is improved by combining dilated convolution with gated convolution. The gating mechanism is used to effectively filter and enhance key information, so that the model can capture important features more accurately when processing short sequences.

[0097] 2. The network construction method of Neural Ordinary Differential Equation is adopted, and the adjoint sensitivity back propagation method is used to reduce memory overhead, build a deeper network, and make the local extracted features more accurate;

[0098] 3. The self-attention mechanism is introduced to identify the dependencies between features in the entire sequence, thereby capturing global information. By combining local and global interactive features, the accuracy and reliability of fault detection can be effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 It is an elevator fault diagnosis model based on short sequence temporal convolutional network.

[0100] Figure 2 This is a comparison chart of the average accuracy of eight models under three label rates.

[0101] Figure 3 This is a comparison chart of the recall rate, precision rate and F1-Score (%) of the four models for six tags. DETAILED DESCRIPTION

[0102] The present invention will be further described below in conjunction with the accompanying drawings.

[0103] Reference Figure 1~Figure 3 , an elevator fault diagnosis method based on a short sequence time convolutional network, the method comprising the following steps:

[0104] The first step is to collect the original vibration signal of the elevator car running acceleration: install an acceleration sensor in the elevator car, and realize remote data acquisition through the TCP / IP communication protocol; use the serial port connection method to connect the sensor, elevator and gateway, and use Modbus and RS485 protocols for data communication;

[0105] The second step is to obtain the original vibration signal. The process is as follows:

[0106] Step (2.1) calculates the time domain characteristic index and adjusts the elevator car vibration acceleration signal by the time domain analysis method. The process is as follows:

[0107] Step (2.1.1) calculates the peak-to-peak value in the vibration signal samples , the formula is as follows,

[0108] ;

[0109] ;

[0110] ;

[0111] in, is the sampling value of the vibration acceleration signal, and max is the sampling value The maximum value, min is the sampling value The minimum value of is the maximum value in the signal samples, is the minimum value in the signal samples;

[0112] The peak-to-peak value is an indicator used to describe the size of the signal variation range within a period. For non-periodic car vibration signals, the peak-to-peak value can be used to measure the amplitude or amplitude variation of the signal, that is, the maximum deviation of the signal.

[0113] Step (2.1.2) calculates the average amplitude of the vibration signal samples RMS , the formula is as follows,

[0114] ;

[0115] ;

[0116] Where N is the number of sampling points, ∑ is the summation operation, which is used in mathematics and statistics to add a series of values. for The absolute value of

[0117] Among them, the average amplitude is used to describe the degree of change of the vibration signal caused by the change of the axis position; the root mean square reflects the average energy intensity and stability of the vibration signal;

[0118] Step (2.1.3) calculates the square root amplitude of the vibration signal samples , the formula is as follows,

[0119] ;

[0120] Among them, the square root amplitude is an indicator used to measure the signal amplitude, which can more accurately reflect the energy distribution and overall strength of the signal;

[0121] Step (2.1.4) adjusts the original vibration signal according to the calculated time domain features , the formula is as follows,

[0122] ;

[0123] in, , They are the target peak-to-peak value, target average amplitude, target RMS value, and target RMS amplitude. is a noise signal used to adjust the vibration signal;

[0124] Step (2.2) segments the vibration signal processed in step (2.1), samples it using a sliding window containing M data points, and then slides the window backward by a number of data points greater than M to intercept the original signal, and finally obtains a valid original vibration signal. ;

[0125] Step (2.3) is as follows Figure 1 As shown in the figure, the signal dimension reduction is achieved through one-dimensional convolution, the convolution kernel is 16, the convolution kernel width is 12, and the step size is 4.

[0126] ;

[0127] in, represents the convolution operation output of the vibration signal corresponding to the i-th feature map, represents the activation function, represents convolution, is a one-dimensional convolution kernel, T is the number of feature maps;

[0128] Among them, one-dimensional convolution is the most widely used method for data dimensionality reduction in convolutional networks;

[0129] Step (2.4) performs maximum pooling on the signal obtained by one-dimensional convolution;

[0130] Among them, the output after maximum pooling is,

[0131] ;

[0132] Among them, z represents the width of the maximum pooling, and then the pooling result is subjected to the dropout operation. Dropout can randomly discard some neurons to prevent the model from overfitting;

[0133] Among them, the role of the maximum pooling layer is to reduce the number of features and extract key features;

[0134] Step (2.5) is as follows Figure 1 As shown in the data preprocessing section, the final processed vibration signal is obtained.

[0135] ;

[0136] The third step is to optimize the TCN model, introduce gated convolution and neural ordinary differential equations into dilated convolution and residual connection to form an improved TCNODE model;

[0137] Among them, Temporal Convolutional Networks (TCN) is a network model proposed by Shaojie Bai, whose architecture is mainly achieved by changing Convolutional Neural Networks (CNN). Dilated convolution expands the coverage of the convolution kernel by introducing a dilation factor without increasing the number of parameters or computational complexity; residual connection makes it easier to optimize the deep neural network structure by introducing residual blocks, which greatly facilitates the application of deep network structures;

[0138] The process of the third step is as follows:

[0139] Step (3.1) The dilated convolution in TCNODE introduces gated convolution, where gated convolution is a component of advanced convolutional neural networks that is mainly responsible for controlling the transmission of information flow; the process is as follows:

[0140] Step (3.1.1) is as follows Figure 1 As shown in the TCNODE section, the present invention uses convolutional layers and loop structures to construct gated convolutions. In gated convolutions, the output of each convolutional layer passes through a gating mechanism that uses a sigmoid activation function to determine the amount of information transferred in each channel. The width of the convolution kernel in the gated convolution layer is , the gating mechanism generates a gating value between 0 and 1 for each channel to dynamically adjust the flow of information, The number of channels is 2, 3, or 6 to cover all cycles of vibration data, and the number of channels is 64, 128, or 256.

[0141] ;

[0142] in, is the gated convolutional layer, and are the weights and biases of the gated convolutional layer, represents the convolution operation, The convolution kernel size is The network parameters when , the sigmoid function is a smooth curve function widely used to map real numbers to the (0,1) interval, and is often used as an activation function in binary classification problems;

[0143] Step (3.1.2) is as follows Figure 1As shown in the TCNODE section, the present invention uses convolutional layers and loop structures to construct dilated convolutions. Dilated convolutions expand the receptive field by increasing the spacing between each unit in the convolution kernel (dilation rate P) without increasing the number of parameters. The dilation rate P is 1, and the convolution kernel width Choose 2, 3, 6, and the number of channels is 64, 128, 256.

[0144] ;

[0145] in, is the dilated convolutional layer, and are the weights and biases of the dilated convolutional layer, represents the convolution operation, The convolution kernel size is The network parameters when , the tanh function is an S-type activation function that maps real numbers to the (-1,1) interval, and is often used in neural networks to centralize data;

[0146] Step (3.1.3) obtains the convolutional layer formula of TCNODE,

[0147] ;

[0148] in, represents element-wise product, is the dilated convolution output matrix obtained in step (3.1.2), is the output vector, is the gated convolution output matrix obtained in step (3.1.1), is the output vector;

[0149] Step (3.2) improves the residual connection method to the Neural Ordinary Differential Equation method, the process is as follows:

[0150] Step (3.2.1) TCN extracts and aggregates temporal information in a non-recursive manner by stacking multiple residual convolution blocks. The local feature L of the sub-block is obtained by the following formula:

[0151] ;

[0152] Among them, is the individually parameterized temporal convolutional layer in step (3.1), Represents a truncation function, taking only The last dimension of elements, The sequence length is The output of the tth layer. The residual output The last dimension of must be truncated to before adding its transformation , because the dilated convolution causes the length of the potential representation to gradually decrease, i.e. , special ,in, is the initial state, r, k , l , R are the dilation factor, convolution kernel size, step size and model receptive field respectively. To ensure that R > T to achieve lossless encoding of all data, the present invention defines that when r=1, , when r>1, ;

[0153] Step (3.2.2) The idea of ​​the Neural ODE is to treat time as the number of layers. By inserting more layers and taking smaller integration steps, we can directly parameterize and approximate the continuous evolution of the underlying state.

[0154] ;

[0155] in, For function The output, For function Parameters;

[0156] Step (3.2.3) converts the propagation depth in TCNODE to Decomposed into integration time and step length The combination of , the magical ordinary differential equation representation of TCNODE can be realized by increasing the variable integration step size dt,

[0157] ;

[0158] Step (3.2.4) for the termination time and the initial state , the present invention is implemented by , a set of parameters Θ is used to characterize the entire continuous time aggregation process, and the local features L of the sub-blocks are extracted.

[0159] ;

[0160] in, The integration process is used to solve the cumulative amount of a function on a specific interval. Figure 1 middle It is a black-box ODE solver used to solve neural ordinary differential equations.

[0161] Step 4: If Figure 1As shown in the Self-attention section, the self-attention mechanism is added to form an improved SA-TCNODE model. The process is as follows:

[0162] Step (4.1) The local feature output sequence of the sub-block obtained in the previous section Through adaptive linear mapping, it is converted into Y length and use it to generate the query vector Q, key vector K and value vector V. Figure 1 As shown, the generation formulas of Q, K, and V are as follows:

[0163] ;

[0164] in, , , It is a learnable weight matrix, which can increase the expressiveness of the model, so that the calculated weight matrix has higher generalization ability;

[0165] Step (4.2) Calculate the attention coefficient based on the Q and K vectors ,

[0166] ;

[0167] in The probability represents the degree of association between two vectors, e is the base of the natural logarithm, ' is the transposition number, and the softmax function is a normalized exponential function in multi-class classification problems;

[0168] Step (4.3) calculates the weights of each sub-block by weighted summing of the attention coefficient value vector V and concatenates the weighted local features to obtain the global feature. The global feature G can be expressed as,

[0169] ;

[0170] Step (4.4) converts the global features The output M is input into the fully connected layer, and finally the output M is input into the softmax function for normalization to obtain the fault diagnosis result.

[0171] ;

[0172] Among them, c is the dimension of the softmax function input, is the probability of the diagnosis object category;

[0173] Step 5: Figure 1 As shown, the adjoint sensitivity back propagation method is introduced to train the model, and the process is as follows:

[0174] Step (5.1) uses the cross entropy loss function to train the model. The calculation formula is as follows:

[0175] ;

[0176] in, represents the actual label, N represents the total number of samples, is the inverse operation of exponential operation, It is a summation operation. Finally, the Adam method is used as the optimizer to optimize the weights and bias terms;

[0177] Step (5.2) loss Network parameters The gradient of can be directly calculated by the adjoint function and the ODE solver,

[0178] ;

[0179] in, is the adjoint matrix.

[0180] The implementation process of the elevator fault diagnosis method based on the short sequence time convolutional network of this embodiment is as follows:

[0181] Step 1: Define the following fault diagnosis model:

[0182] Resnet: It introduces a residual structure, supports the construction of ultra-deep networks, abandons Dropout, and combines batch normalization technology to accelerate the training process of fault diagnosis models.

[0183] LSTM: By introducing a gating mechanism, it can effectively capture long-term dependencies in sequence data and solve the gradient vanishing and gradient exploding problems of traditional RNN in long sequence training.

[0184] GRU: Through a simplified gating mechanism, it can efficiently capture dependencies in sequence data while alleviating the gradient vanishing problem of traditional RNN.

[0185] CAFCL: Based on the feature contrast learning method, it improves the classification ability and robustness of the model by enhancing the consistency of intra-class features and the distinguishability of inter-class features.

[0186] SAE: Stack multiple autoencoders to extract high-level features of the data layer by layer to achieve effective feature representation and dimensionality reduction.

[0187] TFD-DCNN: Improves the analysis and classification capabilities of non-stationary signals by extracting time-frequency domain information.

[0188] SIFT: Based on computer vision algorithms, it detects and describes local feature points in images to achieve scale- and rotation-invariant matching and recognition of images.

[0189] SA-TCNODE: The fault diagnosis method of the present invention integrates the neural ordinary differential equation and the time convolution network to realize local feature extraction, and combines the self-attention mechanism to realize the fusion of global features and local features, and finally realizes the fault diagnosis method of short sequence.

[0190] Step 2: Experimental Dataset

[0191] The experimental data used in the present invention is derived from real-time elevator operation data in sensors remotely acquired using the TCP / IP communication protocol.

[0192] Among them, the number of samples collected for normal vibration of the elevator car is 255,634, the number of samples for excessive vibration in the elevator start-up area is 18,944, the number of samples for excessive vibration in the elevator acceleration area is 18,152, the number of samples for excessive vibration in the elevator uniform speed area is 25,738, the number of samples for excessive vibration in the elevator deceleration area is 20,215, the number of samples for excessive vibration in the elevator stop area is 24,844, and the number of samples for emergency stop failure is 12,846.

[0193] After feature extraction of these data samples and grouping the data sets according to label rates of 5%, 10% and 20%, they can be used as the following experimental data sets.

[0194] Step 3: Define evaluation metrics

[0195] The present invention uses the accuracy rate A to quantitatively analyze the classification performance of the algorithm used.

[0196] ;

[0197] in, represents the number of samples that are diagnosed as positive. Indicates the number of samples that diagnose positive samples as negative. Indicates the number of samples that are diagnosed as positive negative samples. Represents the number of samples that are diagnosed as negative. Accuracy is used to measure the overall ability of the model to correctly diagnose faults.

[0198] Step 4: Analyze and compare the results

[0199] refer to Figure 2 , which shows the accuracy A of eight models (SA-TCNODE, SIFT, TFD-DCNN, SAE, CAFCL, GRU, LSTM, Resnet) when the label rate LR is 5%, 10% and 20% respectively. Figure 2In the data set with label rate LR of 10% and 20%, SA-TCNODE obtained the best values ​​of 0.972 and 0.993, which is better than other comparison methods. The poor performance of other comparison methods can be attributed to two reasons. First, the limited batch size hinders the ability of baseline methods to extract discriminative features from insufficient negative samples. Second, these methods only focus on local information and ignore global information. On the contrary, SA-TCNODE builds a deeper network by combining TCN and NODE, which does not rely on learning potential features from a large number of negative samples, and the self-attention mechanism enables SA-TCNODE to effectively capture important features in local and global data. However, SA-TCNODE is slightly inferior to CAFCL in accuracy at 5% label rate. The reason for this result is that these baseline models use longer samples in the diagnosis process. Among them, CAFCL adopts contrastive learning method to realize semi-supervised fault diagnosis with few labels, which makes the accuracy of the model much higher than other models at low label rate, but it also requires a long sample sequence input for the model, and the accuracy of the model is not significantly improved with the increase of label rate. The sample length of SA-TCNODE is only 150, which achieves high accuracy with a very short sample length, which is of great significance for the real-time diagnosis task of elevator faults.

[0200] This embodiment uses an ablation experiment to prove the effectiveness of the improved fault diagnosis method of the present invention, and the steps are as follows:

[0201] Step 1: Define 4 fault diagnosis models:

[0202] TCN: The fault diagnosis method of the present invention removes the self-attention mechanism module and the neural ordinary differential equation part.

[0203] SA-TCN: The fault diagnosis method of the present invention removes the neural ordinary differential equation part.

[0204] TCNODE: The fault diagnosis method of the present invention removes the self-attention mechanism module.

[0205] SA-TCNODE: The fault diagnosis method of the present invention.

[0206] Step 2: Experimental Dataset

[0207] The experimental data used in this embodiment is the same as that in the previous process, and a label rate of 10% is taken as the following experimental data set.

[0208] Step 3: Define evaluation metrics

[0209] The present invention uses the following classification evaluation indicators to quantitatively analyze the classification performance of the algorithm used. These include recall rate R, precision rate P, and F1-Score,

[0210] ;

[0211] ;

[0212] ;

[0213] in, represents the number of samples that are diagnosed as positive. Indicates the number of samples that diagnose positive samples as negative. Indicates the number of samples that are diagnosed as positive negative samples. It indicates the number of samples that are diagnosed as negative. Recall rate indicates the proportion of all faulty samples that the model can identify. Precision rate measures the accuracy of the model in correctly identifying faulty samples. F1-Score strikes a balance between precision rate and recall rate, providing a comprehensive evaluation of the model classification performance.

[0214] Step 4: Analyze and compare the results

[0215] refer to Figure 3 , which shows the recall rate R, precision rate P, and F1-Score of four models (TCN, SA-TCN, TCNODE and SA-TCNODE). Figure 3 The experimental results show that TCN has the worst overall diagnostic ability, which indicates that many fault samples are misclassified as other fault labels. SA-TCN, which introduces the self-attention mechanism, performs well in fault classification, especially accurately identifying labels 5 and 6, thanks to the advantages of the self-attention mechanism in global feature extraction. TCNODE achieves accurate judgment of more labels by building a deeper network structure, but the effect is still not ideal when processing labels 2 and 3. The SA-TCNODE model outperforms other models in comprehensive performance, especially when processing similar labels. This shows that by combining the self-attention mechanism and the neural ordinary differential equation, TCN can be effectively improved so that it can learn more discriminative features from unlabeled data. In summary, the experimental results verify the effectiveness of the improved method of the present invention and demonstrate the superiority of the SA-TCNODE model in global feature extraction and complex label classification.

[0216] The contents described in the embodiments of this specification are merely enumerations of implementation forms of the inventive concept and are for illustrative purposes only. The protection scope of the present invention should not be considered to be limited to the specific forms described in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be thought of by ordinary technicians in this field based on the inventive concept.

Claims

1. An elevator fault diagnosis method based on short sequence time convolutional network, characterized in that: The method comprises the following steps: The first step is to collect the original vibration signal of the elevator car running acceleration: install an acceleration sensor in the elevator car, and realize remote data acquisition through the TCP / IP communication protocol; use the serial port connection method to connect the acceleration sensor, elevator and gateway, and use Modbus and RS485 protocols for data communication; The second step is to obtain the original vibration signal. The processing process is as follows: Step (2.1) calculates the time domain characteristic index and adjusts the elevator car vibration acceleration signal by using the time domain analysis method; Step (2.2) segments the vibration signal processed in step (2.1), samples it using a sliding window containing M data points, and then slides the window backward by a number of data points greater than M to intercept the original signal, and finally obtains a valid original vibration signal; Step (2.3) reduces the dimensionality of the signal through one-dimensional convolution; Step (2.4) performs maximum pooling on the signal obtained by one-dimensional convolution; Step (2.5) obtains the final processed vibration signal; The third step is to optimize the TCN model, introduce gated convolution and neural ordinary differential equations into dilated convolution and residual connection to form an improved TCNODE model. The process is as follows: Step (3.1) The dilated convolution in TCNODE introduces gated convolution, and the process is as follows: Step (3.1.1) uses convolutional layers and loop structures to construct gated convolutions. In gated convolutions, the output of each convolutional layer passes through a gating mechanism that uses a sigmoid activation function to determine the amount of information transferred in each channel. The width of the convolution kernel in the gated convolution layer is n. G , the gating mechanism generates a gating value between 0 and 1 for each channel to dynamically adjust the flow of information, n G The number of channels is 2, 3, or 6 to cover all cycles of vibration data, and the number of channels is 64, 128, or 256. Among them, F G (V t ) is a gated convolutional layer, and are the weights and biases of the gated convolutional layer, * represents the convolution operation, The convolution kernel size is 1×n G The network parameters when , the sigmoid function is the activation function; Step (3.1.2) uses convolutional layers and loop structures to construct dilated convolution. Dilated convolution expands the receptive field by increasing the spacing between each unit in the convolution kernel, that is, the dilation rate P, without increasing the number of parameters. The dilation rate P is 1, and the convolution kernel width n K Choose 2, 3, 6, and the number of channels is 64, 128, 256. Among them, F K (V t ) is the dilated convolutional layer, and are the weights and biases of the dilated convolutional layer, * denotes the convolution operation, The convolution kernel size is 1×n K The network parameters when , the tanh function is an S-type activation function that maps real numbers to the (-1,1) interval; Step (3.1.3) obtains the convolutional layer formula of TCNODE, TCN(V t ,Θ)=F G (V t )⊙F K (V t )=V G (t)⊙V K (t); Among them, ⊙ represents the element-by-element product, V K (t) = {v K (1),v K (2),…,v K (T)} is the dilated convolution output matrix obtained in step (3.1.2), v K is the output vector, V G (t) = {v G (1),v G (2),…,v G (T)} is the gated convolution output matrix obtained in step (3.1.1), v G is the output vector; Step (3.2) improves the residual connection method to the Neural Ordinary Differential Equation method, and the process is as follows: Step (3.2.1) TCN extracts and aggregates timing information in a non-recursive manner by stacking multiple residual convolution blocks. The local feature L of the sub-block is obtained by the following formula: L t+1 =Γ(V t ,Q t+1 )+TCN(V t ,Θ),t∈{0,…,T-1}; Among them, TCN (V t ,Θ) is the individually parameterized temporal convolutional layer in step (3.1), Γ(·,Q t+1 ) represents the truncation function, taking only L t+1 The last dimension Q in t+1 elements, L t+1 The sequence length is Q t The output of the tth layer, the residual output L t+1 The last dimension of must be truncated to Q before adding its transformation t+1 , because the dilated convolution will cause the length of the potential representation to gradually decrease, that is, Q t+1 =Q t -r t ×(k-1), Q1=R-k+1, where L0 is the initial state, r, k, l, and R are the dilation factor, convolution kernel size, step size, and model receptive field respectively; to ensure that R>T so as to achieve lossless encoding of all data, it is defined that when r=1, R=l(k-1)+1, and when r>1, R=(r t -1)(k-1) / (r-1)+1; Step (3.2.2) The Neural Ordinary Differential Equation is as follows, Where h(t) is the output of function f(t), and θ is the parameter of function f(t); Step (3.2.3) sets the propagation depth k in TCNODE * It is decomposed into a combination of integration time T and step size Δt, namely By adding a variable integration step size dt, the magical ordinary differential equation representation of TCNODE is realized. L k+1 =L k +TCN(L(t),Θ)dt; Step (3.2.4) for the termination time T * and the initial state L0, by setting Δt→0, a set of parameters Θ is used to characterize the entire continuous time aggregation process, and the local features L of the sub-blocks are extracted. in, The integration process is used to solve the cumulative amount of a function on a specific interval. ODESolve is a black-box ODE solver used to solve neural ordinary differential equations. Step 4: Add the self-attention mechanism to form an improved SA-TCNODE model for elevator fault diagnosis. The process is as follows: The local feature output sequence of the sub-block in step (4.1) is converted into Y vectors through adaptive linear mapping, and used to generate the query vector Q, key vector K and value vector V; Step (4.2) calculates the attention coefficient based on the Q and K vectors; Step (4.3) performs weighted summation of the attention coefficient value vector V to calculate the weight of each sub-block and concatenates the weighted local features to obtain the global feature; In step (4.4), the global feature G is input into the fully connected layer to obtain the output M, and finally the output M is input into the softmax function for normalization to obtain the elevator fault diagnosis result.

2. The elevator fault diagnosis method based on short sequence time convolutional network according to claim 1 is characterized in that: The fault diagnosis method further comprises the following steps: Step 5: Introduce the adjoint sensitivity back propagation method to train the model. The process is as follows: Step (5.1) uses the cross entropy loss function to train the model; In step (5.2), the gradient of the loss with respect to the network parameters is directly calculated using the adjoint function and the ODE solver.

3. The elevator fault diagnosis method based on short sequence time convolutional network according to claim 1 or 2, characterized in that: The process of step (2.1) is as follows: Step (2.1.1) calculates the peak value F3 in the vibration signal sample using the following formula: F1=max(v i ); <h2 style=";text-align:left;direction:ltr">F2 = min(v<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> ); F3 = F1-F2; Among them, v i is the sampling value of the vibration acceleration signal, and max is the sampling value v i The maximum value, min is the sampling value v i The minimum value of , F1 is the maximum value in the signal sample, and F2 is the minimum value in the signal sample; Step (2.1.2) calculates the average amplitude F4 and root mean square F5 in the vibration signal samples, the formula is as follows, Where N is the number of sampling points, ∑ is the summation operation, which is used in mathematics and statistics to represent the addition of a series of values, |v i |v i The absolute value of Step (2.1.3) calculates the square root amplitude F6 in the vibration signal sample, the formula is as follows, Step (2.1.4) adjusts the original vibration signal V(τ) according to the calculated time domain characteristics. The formula is as follows: in, φ and λ are the target peak-to-peak value, target average amplitude, target RMS value, and target RMS amplitude, respectively. n(τ) is the noise signal used to adjust the vibration signal.

4. The elevator fault diagnosis method based on short sequence time convolutional network according to claim 1 or 2, characterized in that: The step (2.2) obtains the effective original vibration signal V(τ)={v1,v2,…,v N }; In step (2.3), the convolution kernel is 16, the convolution kernel width is 12, and the step length is 4. Among them, V i (τ) represents the convolution operation output of the vibration signal corresponding to the i-th feature map, f(·) represents the activation function, * represents convolution, W is the one-dimensional convolution kernel, and T is the number of feature maps; In the step (2.4), the output after maximum pooling is, V i =max{V i,1 ,V i,2 ,…,V i,z },i=1,2,…,T; Among them, z represents the width of the maximum pooling, and then the pooling result is dropped out; In the step (2.5), the final processed vibration signal is obtained. V t =V1,V2,…,V T 。 5. The elevator fault diagnosis method based on short sequence time convolutional network according to claim 1 or 2, characterized in that: In the step (4.1), the local feature output sequence of the sub-block is L = [L 1 ,L 2 ,…,L Y ] is transformed into Y pieces of length d through adaptive linear mapping. K The vector is used to generate the query vector Q, key vector K and value vector V. The generation formula of Q, K and V is as follows: Among them, W Q , W K , W V is the learnable weight matrix; In the step (4.2), the attention coefficient α is calculated based on the Q and K vectors. where α=[α1,α2,…,α n ] uses probability to represent the degree of association between two vectors, e is the base of the natural logarithm, ' is the transposition number, and the softmax function is a normalized exponential function in multi-class classification problems; in the step (4.3), the global feature G is expressed as, In step (4.4), normalization is expressed as, Among them, c is the dimension of the softmax function input, y i is the diagnosis object category probability.

6. The elevator fault diagnosis method based on short sequence time convolutional network according to claim 2 is characterized in that: In step (5.1), the cross entropy loss function is used to train the model, and the calculation formula is as follows: Among them, l i represents the actual label, N represents the total number of samples, log is the inverse operation of the exponential operation, ∑() is the summation operation, and finally the Adam method is used as the optimizer to optimize the weights and bias terms; In the step (5.2), the loss The gradient of the network parameter θ is directly calculated by the adjoint function and the ODE solver. in, is the adjoint matrix.

Images