Steel wire rope wear detection and classification method based on vibration signals, medium and equipment

Abstract

The invention provides a steel wire rope wear detection and classification method based on vibration signals, a medium and equipment, and relates to the field of fault diagnosis. The core technology is that vibration signals of contact between a steel wire rope and a pulley in different abrasion states when a bus door is opened are collected; characteristic values of the vibration signals are extracted, characteristic subsets meeting preset conditions are screened out from the characteristic values to serve as final characteristic values of steel wire rope abrasion diagnosis, a multi-channel characteristic matrix is formed, and a test set and a training set are divided; and constructing a CNN-LSTM feature coupling model, extracting spatial features by using CNN, capturing a time sequence rule by using LSTM, iteratively optimizing hyper-parameters in combination with IWOA, realizing high-precision wear classification, and outputting the predicted wear state of the steel wire rope. According to the method, human subjective factor interference can be avoided; meanwhile, the strong dependence of materials analysis, electromagnetic analysis and computer vision schemes on the working environment is eliminated, the judgment deviation caused by environmental interference is reduced, and the stability and reliability of the detection result are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis, and in particular to a method, medium, and equipment for detecting and classifying wire rope wear based on vibration signals. Background Technology

[0002] Currently, traditional methods for detecting wear on the steel wire ropes of bus gates mostly rely on simple mechanical measurements, which are easily affected by the experience and subjective factors of the inspectors. A more advanced approach is to conduct materials science or electromagnetic analysis on the material properties of the steel wire rope, or to use computer vision to identify the degree of wear and thus predict the wear condition of the steel wire rope.

[0003] In actual operation, the assessment of wire rope wear conditions by methods such as materials science analysis, electromagnetic analysis, and computer vision is subject to various working environments and conditions, resulting in significant deviations. Summary of the Invention

[0004] Purpose of the invention: To overcome the problems existing in the prior art, this invention proposes a method, medium, and device for detecting and classifying wire rope wear based on vibration signals. The method involves collecting vibration signals generated when the wire rope of a bus door drive device rubs against a pulley using an accelerometer. Spatial features are extracted using a convolutional neural network (CNN), and temporal dynamic patterns are captured using a long short-term memory network (LSTM). A CNN-LSTM feature coupling model is constructed, and the hyperparameters of the CNN-LSTM are iterated using IWOA to build the optimal feature coupling model, thereby achieving high-precision prediction of the wire rope wear state.

[0005] To achieve the above objectives, this invention proposes a method for classifying and detecting wire rope wear based on vibration signals, comprising the following steps: Vibration signals of the steel wire rope in contact with the pulley under different wear conditions when the bus door is opened were collected; Extract the feature values ​​of the vibration signal and perform normalization processing; The feature subset that meets the predetermined conditions is selected from all feature values ​​as the final feature values ​​for wire rope wear diagnosis, and a multi-channel feature matrix is ​​formed to divide the test set and training set. Construct a CNN-LSTM network, train it using the training set, and use the improved whale algorithm to determine the optimal hyperparameters of the CNN-LSTM network to obtain the IWOA-CNN-LSTM network; The test set is input into the IWOA-CNN-LSTM network, which outputs the predicted wear state of the wire rope.

[0006] As a preferred approach, vibration signals from the contact between the wire rope and pulley under different wear conditions during the opening of the bus door are collected, specifically including: Accelerometers are installed at the center, eccentric, and edge positions of the pulley disc. Each accelerometer serves as a data acquisition channel, collecting vibration signals generated by the wire rope in contact with the pulley under four different wear conditions: no wear, slight wear, moderate wear, and severe wear. These signals serve as the original wear signals of the wire rope under different wear conditions.

[0007] As a preferred approach, several time-domain indices and entropy values ​​are extracted from the original wear signal as original feature values.

[0008] As a preferred approach, the original wear signals from all three channels are one-dimensional sampling sequences, with each channel containing N sampling points, denoted as... Where m=1,2,3; n=1,2,...,N, This represents the Nth time-domain index or entropy value collected by the m-th accelerometer.

[0009] As a preferred option, the wear-resistant raw signals of the three channels are... , , The time-domain metrics and entropy values ​​are calculated independently, and the time-domain metrics include peak-to-peak values. Kurtosis K, Root Mean Square (RMS), Waveform Factor The entropy values ​​include energy entropy EE, approximate entropy ApEn, sample entropy SampEn, and dispersion entropy DE.

[0010] As a preferred embodiment, the composition of the multi-channel feature matrix specifically includes: The time-domain metrics and entropy values ​​of each channel are concatenated in a fixed order: "peak-to-peak value → kurtosis → root mean square → waveform factor → energy entropy → approximate entropy → sample entropy → dispersion entropy," forming three sets of one-dimensional feature vectors. Channel 1: ; Channel 2: ; Channel 3: ; The three sets of one-dimensional feature vectors are arranged according to the rule of "one row corresponds to one channel and one column corresponds to one feature" to form a two-dimensional structured matrix. The matrix form is as follows: .

[0011] As a preferred embodiment, in the CNN-LSTM network: The input layer receives data with dimensions [8,3,1] after sequence folding, where 8 is the number of features, 3 is the number of channels, and 1 is the single temporal step size; The CNN module contains two cascaded convolutional units. Each cascaded convolutional unit is configured with 10 convolutional kernels, ReLU activation function, He initialization and batch normalization. The first convolutional unit uses a [2,3]-dimensional convolutional kernel to adapt to 8×3 input. The second convolutional unit uses a [1,3]-dimensional convolutional kernel, followed by a [1,2]-dimensional max pooling layer with a stride of 1. After sequence unfolding and flattening, it is connected to the LSTM module. The LSTM module is a two-layer stacked structure, with batch normalization and Xavier initialization enabled in both layers; a Dropout layer with a ratio of 0.4 is connected after the output of the second layer to suppress overfitting. The output layer maps the LSTM module output to 4 dimensions through a fully connected layer, corresponding to four diagnostic types. The output probability distribution is processed by the Softmax function, and cross-entropy is used as the loss function.

[0012] As a preferred approach, the optimal hyperparameters of the CNN-LSTM network are determined using the improved Whale Algorithm IWOA, resulting in the IWOA-CNN-LSTM network. In the improved Whale Algorithm IWOA: Introducing cosine factor Introducing adaptive weights Where t is the current iteration number, Indicates the maximum number of iterations; By continuously adjusting the whale's position and speed using the following formula, the value of the objective function is made to continuously approach the optimal solution:

[0013]

[0014] In the formula, The position vector representing the current optimal individual; r is a random value in the range [0,1]. represents the distance vector from the individual whale to the current best solution; l is any value in the interval [-1,1]; b is the spiral shape parameter constant.

[0015] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program that enables a computer to execute the above-described method for classifying and detecting wire rope wear based on vibration signals.

[0016] Furthermore, the present invention also proposes an electronic device, characterized in that it includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the above-described method for classifying and detecting wire rope wear based on vibration signals.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: (1) Vibration signals are collected by an accelerometer to avoid interference from human subjective factors; at the same time, the strong dependence of material analysis, electromagnetic analysis and computer vision solutions on the working environment is eliminated, the judgment deviation caused by environmental interference is reduced, and the stability and reliability of the detection results are significantly improved.

[0018] (2) The entire testing process is based on vibration signal acquisition and intelligent algorithm analysis. It does not require complex material disassembly or special testing environment. It is convenient to operate and responds quickly. The constructed model can be directly applied to the real-time detection of the steel wire rope of the bus door after data training and verification. It can quickly output the wear status diagnosis results, provide timely basis for equipment maintenance, and reduce the risk of failure and maintenance costs.

[0019] (3) The core logic of this technical solution can be transferred to wire rope wear detection in other scenarios. It can be adapted to the detection needs of different devices by simply adjusting the signal acquisition parameters and model training data according to the actual application scenario. It has a wide range of applications. At the same time, it adopts a standardized algorithm framework and data processing flow, which facilitates subsequent technology upgrades and system integration. Attached Figure Description

[0020] Figure 1 This is a flowchart of the wire rope wear detection and classification method based on vibration signals according to the present invention.

[0021] Figure 2 This is a diagram of the actual vibration signal collected by the bus door in this invention.

[0022] Figure 3 This invention presents three types of wire rope wear patterns.

[0023] Figure 4 The experimental performance curves for distinguishing different wear states of steel wire ropes in this invention are shown in the overall accuracy diagram.

[0024] Figure 5 This invention provides a confusion matrix diagram for different models of wire ropes under different wear conditions.

[0025] Figure 6 This is a diagram showing the actual arrangement of the three sets of acceleration sensors in this invention. Detailed Implementation

[0026] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0027] In one embodiment, the present invention employs a wire rope wear detection and classification method based on vibration signals, such as... Figure 1 As shown, the steps for fault feature extraction are as follows: Step 1: The wire rope vibration signal acquisition device collects vibration signals when the wire rope rotates around the pulley and contacts the pulley during the opening and closing of the bus door. A training set is constructed by analyzing the time-domain and frequency-domain characteristics of the wire rope vibration signals under different wear conditions to detect the wear state of the wire rope. Step 2: Use a wire rope vibration signal acquisition device to collect the vibration signal of the wire rope under no-wear conditions. The sampling frequency is 4000 Hz. The time-domain plot of the vibration signal collected by the sensor is shown below. Figure 4 As shown in the figure. The total number of data points collected is 240, and the data length is divided into four categories: no wear, slight wear, moderate wear, and heavy wear. The wear diagrams of the wire ropes with slight wear, moderate wear, and heavy wear are shown in the figure. Figure 2 As shown. Groups of sixty from each class are synthesized and saved in a single .mat file, forming the overall dataset. The dataset is then divided into training and validation sets in an 8:2 ratio. Step 3: Perform time and frequency domain analysis on the acquired signal. The time-domain indices used in this paper include peak-to-peak value, rectified average value, kurtosis, skewness, root mean square (RMS), waveform factor, peak factor, impulse factor, and margin factor. The frequency-domain entropy values ​​include power spectral entropy, energy entropy, approximate entropy, sample entropy, fuzzy entropy, permutation entropy, and dispersion entropy. The above time and frequency domain signal characteristic values ​​were calculated using MATLAB on the acquired slightly worn signal. Step 4: Select the optimal time and frequency domain metrics based on the scores. The comprehensive scores show that the features with high combined scores for peak-to-peak value, kurtosis, root mean square, waveform factor, energy entropy, approximate entropy, sample entropy, and dispersion entropy contribute significantly to the model's prediction and should be retained. Step 5: IWOA Initialization. Initial whale population size. =30, maximum number of iterations =100; Step 6: Set the optimization range for each hyperparameter in the hybrid model; Step 7: Enable hyperparameter optimization. At the end of each iteration, IWOA outputs the currently found hyperparameters; then it is input into CNN-LSTM to build a temporary model; the MSE between the output value of the temporary model and the actual value is calculated, and this is used as the fitness value of the population, which is the optimization target of IWOA; Step 8: Update population location and related parameters; Step 9: Accuracy Comparison. Calculate the accuracy of the new iteration and compare it with the accuracy after the previous iteration. If it is less than the previous one, it indicates that the current hyperparameters are the optimal hyperparameters, and the CNN-LSTM model is updated accordingly; otherwise, no update is made, and the hyperparameters obtained from the previous optimization remain unchanged. Step 10: Iteration Termination Check. Determine if the iteration termination condition has been met. If so, the current output is the overall optimal hyperparameter, which is used to update the model and train the final model until the accuracy of the wire rope wear state prediction is output. If not, return to Step 8 to continue iterating.

[0028] In Step 1 above, a physical diagram of the wire rope vibration signal acquisition device is shown below. Figure 6 Three accelerometers were placed on the side of the pulley, approximately at the center of the pulley disc, and at the center and edge of the pulley, respectively, to collect vibration signals generated by the contact between the wire rope and the pulley at four different wear levels. These signals served as the four different raw wear signals for the wire rope wear levels.

[0029] The collected data were classified into four datasets (no wear, light wear, moderate wear, and heavy wear), and feature analysis was performed on the extracted signal data.

[0030] In Step 2 above, time-domain indicators and entropy values ​​(peak-to-peak value, rectified average value, kurtosis, skewness, root mean square, waveform factor, peak factor, impulse factor, margin factor, power spectral entropy, energy entropy, approximate entropy, sample entropy index, fuzzy entropy, permutation entropy, and dispersion entropy) are extracted. The mRMR is used to rank the 16 feature values ​​based on their correlation, selecting feature values ​​with low redundancy and high correlation as input to the neural network. After further filtering using mRMR, peak-to-peak value, kurtosis, root mean square, waveform factor, energy entropy, approximate entropy, sample entropy, and dispersion entropy are selected as the final feature values ​​for steel wire rope wear diagnosis based on vibration signals. The specific calculation results are shown in Table 1 below.

[0031] Table 1. Time-frequency domain feature value category scores

[0032] In Step 4 above, after mRMR filtering of feature values, peak-to-peak value, kurtosis, root mean square (RMS), waveform factor, energy entropy, approximate entropy, sample entropy, and dispersion entropy are selected to form a multi-channel feature matrix. This technique extracts four time-domain indicators (peak-to-peak value, kurtosis, RMS, waveform factor) and four entropy values ​​(energy entropy, approximate entropy, sample entropy, and dispersion entropy) for each of the three channels of the original time-series signal. After being aligned in a unified order, they are integrated into a 3×8 dimensional multi-channel feature fusion matrix, which retains the integrity of single-channel features while reflecting cross-channel differences. This provides structured and highly discriminative input for subsequent data analysis tasks. The steps of the composition process are as follows: 1. Data Background Data source: The data acquisition array consists of three accelerometers deployed at different monitoring points, which collect continuous temporal vibration signals of the steel wire rope in contact with the pulley. Each sensor corresponds to one independent channel.

[0033] Data format: The raw signals of the three channels are all one-dimensional sampling sequences, with each channel containing N sampling points, denoted as... (Where m=1,2,3; n=1,2,...,N), three channels are collected synchronously to ensure data timing consistency.

[0034] Raw signals from 3 channels The following features are calculated independently, with all channels strictly following the same feature order to ensure dimensional alignment during subsequent fusion.

[0035] 2. Feature Data Time-domain indicators Peak-to-peak value ( : The difference between the maximum and minimum values ​​of a signal, reflecting the maximum fluctuation range of the signal.

[0036] Kurtosis (K): Describes the steepness of the probability distribution of a signal and is sensitive to impulsive anomalies (such as fault pulses).

[0037] Root mean square (RMS): The square root of the mean of the sum of squares of a signal, representing the effective energy level of the signal.

[0038] Waveform factor ( : The ratio of the root mean square to the rectified average (arithmetic mean of the absolute value of the signal), reflecting the degree of distortion of the signal waveform.

[0039] Entropy characteristics Energy entropy (EE): Calculated based on the energy distribution of a signal. The larger the value, the more dispersed the energy distribution and the more irregular the signal.

[0040] Approximate entropy (ApEn): measures the regularity of signal timing; the larger the value, the higher the signal complexity and the lower the predictability.

[0041] Sample Entropy (SampEn): Optimizes the small sample bias problem of approximate entropy, and more stably reflects the inherent complexity of the signal.

[0042] Dispersion entropy (DE): Combining amplitude discretization and arrangement pattern, it captures both the nonlinear characteristics and amplitude distribution of the signal, and has strong noise resistance.

[0043] 3. Data fusion Single-channel feature vector construction: The four time-domain metrics and four entropy values ​​of each channel are concatenated in a fixed order of "peak-to-peak value → kurtosis → root mean square → waveform factor → energy entropy → approximate entropy → sample entropy → dispersion entropy" to form three one-dimensional feature vectors: Channel 1:

[0044] Channel 2:

[0045] Channel 3:

[0046] Cross-channel feature alignment: The feature vectors of the three channels are all 8-dimensional, and the feature types corresponding to each dimension are completely consistent, avoiding feature confusion during fusion.

[0047] Fusion matrix integration: The three feature vectors are arranged according to the rule of "one row corresponds to one channel and one column corresponds to one feature" to form a two-dimensional structured matrix.

[0048] 4. Specific representation of the channel fusion matrix Matrix dimension definition The fusion matrix is ​​3×8 dimensional (3 rows correspond to 3 channels, 8 columns correspond to 8 features), denoted as ( It has a compact structure and a clear physical meaning.

[0049] Mathematical expression (1) Rows and Columns and the Meaning of Elements Row index (1~3): corresponding to the 1st, 2nd and 3rd data channels (accelerometer 1, accelerometer 2, accelerometer 3), respectively. Each row is a complete feature set of a channel.

[0050] Column indices (1-8): These correspond to 8 fixed features, with the following mapping relationships: Column 1 = Peak-to-peak value | Column 2 = Kurtosis | Column 3 = Root mean square | Column 4 = Waveform factor | Column 5 = Energy entropy | Column 6 = Approximate entropy | Column 7 = Sample entropy | Column 8 = Dispersion entropy element The j-th eigenvalue of the i-th channel (e.g.) ) represents the waveform factor of the second channel, ( () represents the approximate entropy of the 3rd channel.

[0051] This article uses matrix examples. (2) The above matrix is ​​used as the input to the subsequent neural network.

[0052] The mRMR algorithm yields the following optimal time-domain and entropy indices: This method is based on the principles of information theory. It evaluates the importance of features by calculating the correlation and redundancy between features, and then selects the most representative subset of features.

[0053] The definition of mutual information is as follows: (3) in: , These are random variables , The probability density, yes , The joint probability density, express and Mutual information between them.

[0054] The mRMR algorithm uses mutual information theory to calculate the redundancy between features and the correlation between features and input samples. (4) (5) In the formula: This represents the correlation between samples and feature categories in the mRMR algorithm. Indicates redundancy between features Represents the feature set, Indicates the number of features. Indicates the sample category. and They represent feature sets respectively The elements in the mRMR (modular redundancy index) are used to find the optimal features using an incremental search method. A smaller difference in mutual information between relevance and redundancy indicates a more important feature.

[0055] (6) The comprehensive properties of all time-domain and entropy indices are calculated using formulas, and the optimal indices are selected. Time- and frequency-domain analysis is performed on the acquired signals. The time-domain indices used in this paper are peak-to-peak value, rectified average value, kurtosis, skewness, root mean square (RMS), waveform factor, peak factor, impulse factor, and margin factor, and the frequency-domain entropy values ​​are power spectral entropy, energy entropy, approximate entropy, sample entropy, fuzzy entropy, permutation entropy, and dispersion entropy. The above time- and frequency-domain signal characteristic values ​​are calculated using MATLAB on the acquired slightly worn signals. The characteristic values ​​suitable for this signal after mRMR calculation and selection are peak-to-peak value, kurtosis, RMS, waveform factor, energy entropy, approximate entropy, sample entropy, and dispersion entropy.

[0056] In Step 9 above, the IWOA initialization hyperparameters are as follows: In WOA (Whale of Origin and Analysis), the variables for solving the problem are considered as the whale's position and speed. By continuously adjusting the whale's position and speed, the objective function's value is made to continuously approach the optimal solution. WOA includes the following three mechanisms: Exploration mechanism: (7) (8) In the formula, t is the current iteration number. The position vector representing an individual whale. The position vector representing the current optimal individual; ; r is a random value in the range [0,1]. This indicates the maximum number of iterations.

[0057] Predation mechanism: (9) (10) In the formula, represents the distance vector from the individual whale to the current best solution; l is any value in the interval [-1,1], and b is the spiral shape parameter constant, which is set to 1.

[0058] Oscillation mechanism: when the coefficient When, it can be expressed as: (11) (12) In the formula, It is the position vector of a randomly selected individual. When At times, whale pods will choose to conduct more precise searches within a localized area.

[0059] Introduce a cosine factor c to replace m, as shown in the equation: (13) In WOA, adaptive weights are also introduced. Control the impact of the current best individual on the whale pod's position update. The definition of is: (14) The formula is then updated to: (15) (16) The optimal hyperparameters for the CNN-LSTM algorithm are obtained through iterative analysis using IWOA.

[0060] In Step 10 above, the CNN-LSTM algorithm extracts features as follows: First, input the optimal hyperparameters of CNN-LSTM obtained by IWOA iteration to complete the setting of the optimal hyperparameters of CNN-LSTM. Then, use the mRMR algorithm to select the optimal time-domain and frequency-domain indices as input.

[0061] The output of one-dimensional convolution for feature extraction of time series is: (17) In the formula, Features extracted; It is the sigmoid activation function; This is the weight matrix; It is a time series; This is the bias vector. Its specific calculation formula is shown below: (18) (19) (20) (twenty one) (twenty two) (twenty three) In the formula, , , , This is the weight matrix; , , , This is the corresponding bias vector; It is the hyperbolic tangent function; This is matrix multiplication; This is the output from the previous time step; Values ​​representing the degree of retention; The memory state of the previous moment; Add a degree value to the current state; This is an intermediate state; The current state; the output level value; This is the output for the current moment; This is the input for the current moment.

[0062] Based on this, the parameters of the constructed CNN-LSTM hybrid model are set as follows: The input layer receives 10 temporal indicators, transformed from 480 data points, and divided into training and validation sets in an 8:2 ratio; the dimension is [10,1,1], adapted to the CNN through sequence folding; the CNN contains two convolutional units with 10 [2,1] and [1,1] convolutional kernels, with batch normalization and ReLU, He initialization, followed by a [1,3] pooling layer, which is then connected to the LSTM after sequence unfolding and flattening; the LSTM is a 2-layer stacked structure, with the number of neurons optimized by IWOA, with an optimization upper limit of 128, batch normalization and Xavier initialization enabled, followed by dropout with a ratio of 0.4; the output layer is mapped to 4 dimensions through a fully connected layer to correspond to the four diagnostic types; the output probability is obtained through softmax, with cross-entropy as the loss. During training, the initial learning rate, the number of LSTM neurons, the number of training iterations, and the batch size are co-optimized by IWOA: the population size is set to 15, the iterations are 400, and the parameter boundaries are reasonably set; Next, the proposed IWOA-CNN-LSTM and other deep learning models used in this experiment were trained and validated using the same dataset. 80% of the dataset was used as the training set, and 20% as the test set. In the proposed mRMR feature selection IWOA-CNN-LSTM, the feature values ​​are optimized and extracted using mRMR to obtain a feature fusion matrix. This feature fusion matrix is ​​then input into the IWOA-CNN-LSTM model for automatic identification of wire rope wear types. To demonstrate the superiority of mRMR-IWOA-CNN-LSTM, three methods are introduced to analyze the feature data: PSO-CNN-LSTM, WOA-CNN-LSTM, and IWOA-CNN-LSTM. Figure 4 The training processes for four methods are presented. Results show that after 500 training cycles, the training accuracy of the experimental method mRMR-IWOA-CNN-LSTM used in this invention is 97.92%, which is higher than the other three methods (PSO-CNN-LSTM 95.83, WOA-CNN-LSTM 91.67, and IWOA-CNN-LSTM 95.83).

[0063] In another embodiment, the present invention provides a computer-readable storage medium storing a computer program that causes a computer to execute the bus door wire rope fault diagnosis method of the foregoing embodiment.

[0064] In another embodiment, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the bus door wire rope fault diagnosis method of the aforementioned embodiment.

[0065] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0066] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0067] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for classifying and detecting wire rope wear based on vibration signals, characterized in that, Includes the following steps: Vibration signals of the steel wire rope in contact with the pulley under different wear conditions when the bus door is opened were collected; Extract the feature values ​​of the vibration signal and perform normalization processing; The feature subset that meets the predetermined conditions is selected from all feature values ​​as the final feature values ​​for wire rope wear diagnosis, and a multi-channel feature matrix is ​​formed to divide the test set and training set. Construct a CNN-LSTM network, train it using the training set, and use the improved whale algorithm to determine the optimal hyperparameters of the CNN-LSTM network to obtain the IWOA-CNN-LSTM network; The test set is input into the IWOA-CNN-LSTM network, which outputs the predicted wear state of the wire rope.

2. The method for classifying and detecting wire rope wear based on vibration signals according to claim 1, characterized in that, Vibration signals of the steel wire rope contacting the pulley under different wear conditions during the opening of the bus door were collected, specifically including: Accelerometers are installed at the center, eccentric, and edge positions of the pulley disc. Each accelerometer serves as a data acquisition channel, collecting vibration signals generated by the wire rope in contact with the pulley under four different wear conditions: no wear, slight wear, moderate wear, and severe wear. These signals serve as the original wear signals of the wire rope under different wear conditions.

3. The method for classifying and detecting wire rope wear based on vibration signals according to claim 2, characterized in that, Several time-domain indices and entropy values ​​are extracted from the original wear signal as original feature values.

4. A method for classifying and detecting wire rope wear based on vibration signals according to claim 2, characterized in that, The original wear signals from all three channels are one-dimensional sampling sequences, with each channel containing N sampling points, denoted as... Where m=1,2,3; n=1,2,...,N, This represents the Nth time-domain index or entropy value collected by the m-th accelerometer.

5. A method for classifying and detecting wire rope wear based on vibration signals according to claim 4, characterized in that, Raw wear signals for the three channels , , The time-domain metrics and entropy values ​​are calculated independently, and the time-domain metrics include peak-to-peak values. Kurtosis K, Root Mean Square (RMS), Waveform Factor The entropy values ​​include energy entropy EE, approximate entropy ApEn, sample entropy SampEn, and dispersion entropy DE.

6. The method for classifying and detecting wire rope wear based on vibration signals according to claim 5, characterized in that, The multi-channel feature matrix specifically includes: The time-domain metrics and entropy values ​​of each channel are concatenated in a fixed order: "peak-to-peak value → kurtosis → root mean square → waveform factor → energy entropy → approximate entropy → sample entropy → dispersion entropy," forming three sets of one-dimensional feature vectors. Channel 1: ; Channel 2: ; Channel 3: ; The three sets of one-dimensional feature vectors are arranged according to the rule of "one row corresponds to one channel, and one column corresponds to one feature" to form a two-dimensional structured matrix. The matrix form is as follows: 。 7. The method for classifying and detecting wire rope wear based on vibration signals according to claim 1, characterized in that, In the CNN-LSTM network: The input layer receives data with dimensions [8,3,1] after sequence folding, where 8 is the number of features, 3 is the number of channels, and 1 is the single temporal step size; The CNN module contains two cascaded convolutional units. Each cascaded convolutional unit is configured with 10 convolutional kernels, ReLU activation function, He initialization and batch normalization. The first convolutional unit uses a [2,3]-dimensional convolutional kernel to adapt to 8×3 input. The second convolutional unit uses a [1,3]-dimensional convolutional kernel, followed by a [1,2]-dimensional max pooling layer with a stride of 1. After sequence unfolding and flattening, it is connected to the LSTM module. The LSTM module is a two-layer stacked structure, with batch normalization and Xavier initialization enabled on both layers; The second layer output is followed by a Dropout layer with a ratio of 0.4 to suppress overfitting; The output layer maps the LSTM module output to 4 dimensions through a fully connected layer, corresponding to four diagnostic types. The output probability distribution is obtained through the Softmax function, and cross-entropy is used as the loss function.

8. The method for classifying and detecting wire rope wear based on vibration signals according to claim 7, characterized in that, The optimal hyperparameters of the CNN-LSTM network are determined using the improved Whale Algorithm IWOA, resulting in the IWOA-CNN-LSTM network. In the improved Whale Algorithm IWOA: Introducing the cosine factor Introducing adaptive weights Where t is the current iteration number, Indicates the maximum number of iterations; By continuously adjusting the whale's position and speed using the following formula, the value of the objective function is made to continuously approach the optimal solution: In the formula, The position vector representing the current optimal individual; r is a random value in the range [0,1]; D represents the distance between the whale and the optimal solution; represents the distance vector from the individual whale to the current best solution; l is any value in the interval [-1,1]; b is the spiral shape parameter constant.

9. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to execute the wire rope wear detection and classification method based on vibration signals as described in any one of claims 1 to 8.

10. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the wire rope wear detection and classification method based on vibration signals as described in any one of claims 1 to 8.

Images