A gearbox abnormality detection method based on a running-in bench

Abstract

The present application belongs to the field of mechanical fault diagnosis, and discloses a gearbox abnormality detection method based on a running-in platform, comprising the following steps: step 1, collecting data and transmitting, the data including vibration signals, rotation speed signals, pressure signals and torque signals; step 2, analyzing the collected data; step 3, screening and classifying the data under multiple working conditions; step 4, extracting features from the data; and step 5, performing abnormality detection under multiple working conditions. The present application realizes the full-process automation of data collection, transmission, preprocessing, feature extraction and abnormality detection. Through the Cross-Attention mechanism and LSTM and CNN neural networks, the present application realizes the effective fusion of multi-sensor data and enhances the fault feature extraction capability.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis, specifically to a method for detecting abnormalities in gearboxes based on a break-in bench. Background Technology

[0002] Construction machinery is widely used in various industries such as building construction, mining, port transportation, and road construction, encompassing a variety of equipment including excavators, bulldozers, loaders, cranes, and pavers. With the accelerated pace of global infrastructure construction, the construction machinery industry is placing higher demands on the performance, reliability, and intelligence of its equipment. However, because construction machinery operates in complex environments (such as high temperature, high humidity, dust, and heavy loads) for extended periods, its mechanical transmission systems are susceptible to severe impacts and overloads, leading to wear, fatigue, and aging of key components, ultimately causing equipment failure.

[0003] To improve equipment lifespan and safety, the construction machinery industry has gradually introduced intelligent monitoring and predictive maintenance technologies in recent years. These technologies combine intelligent detection, 5G communication, edge computing, and artificial intelligence to achieve real-time monitoring and health management of equipment status. Among these, gearboxes, as core components of power transmission in construction machinery, have become an important research direction for intelligent operation and maintenance, particularly in fault monitoring and anomaly detection.

[0004] Gearboxes are key components in the power systems of construction machinery, responsible for transmitting and regulating power to meet load demands under different operating conditions. During long-term operation, gearboxes are subjected to factors such as high-frequency vibration, complex loads, impact torques, and lubricant deterioration, leading to malfunctions such as tooth surface wear, gear fracture, bearing fatigue, and poor lubrication. These malfunctions can cause equipment performance degradation, increased energy consumption, and even serious safety accidents. Therefore, predictive monitoring and intelligent fault diagnosis of gearboxes are crucial for the safety and stability of construction machinery. Summary of the Invention

[0005] The purpose of this invention is to provide a gearbox anomaly detection method based on a break-in bench, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for detecting gearbox anomalies based on a break-in bench includes:

[0008] Step 1: Collect and transmit data, including vibration signals, rotational speed signals, pressure signals, and torque signals;

[0009] Step 2: Analyze the collected data;

[0010] Step 3: Filter and classify the data according to various operating conditions;

[0011] Step 4: Extract features from the data;

[0012] Step 5: Perform multi-condition anomaly detection.

[0013] Further, step 1 includes:

[0014] First, the collected data is aggregated via a CAN bus. The aggregated data is then compressed and transmitted remotely via a wireless network. During data transmission, data packets are sent in groups of m, with each packet lasting n minutes.

[0015] Further, step 2 includes:

[0016] The server receives data transmitted from the break-in bench in real time and monitors the data packets at preset reception intervals to determine whether a complete set of data has been successfully received. After successfully receiving the complete data, the server automatically performs data decompression and extracts vibration signals, pressure signals, speed signals, and torque signals from the compressed data packets. The decompressed data is automatically sorted and synchronized according to timestamps to ensure that different types of sensor signals remain consistent in the time dimension.

[0017] Further, step 3 includes:

[0018] Wavelet denoising is applied to the vibration signal. Environmental noise and sensor errors are removed through multi-scale analysis. The denoising effect on the vibration signal is expressed by the following formula:

[0019]

[0020] in The wavelet coefficients are processed by soft thresholding, sgn(·) is the sign function, which indicates the positive or negative sign of the input value, W(j,k) represents the k-th coefficient in the j-th level wavelet decomposition, and λ is the set threshold.

[0021] In terms of data classification, a multimodal LSTM neural network is used for feature extraction and time-series modeling of vibration signals, rotational speed signals, pressure signals, and torque signals. Let X1 be the vibration signal, X2 be the rotational speed signal, X3 be the pressure signal, and X4 be the torque signal. The features of each signal are extracted using the LSTM neural network, and the formula can be expressed as:

[0022] h vib =LSTM(X1),h rpm =LSTM(X2)

[0023] h pres =LSTM(X3),h tor =LSTM(X4)

[0024] Where X1, X2, X3, and X4 represent four original time-series signals: vibration, rotational speed, pressure, and torque, respectively; LSTM(·) represents a Long Short-Term Memory network used to extract temporal features; h vib ,h rpm ,h pres ,h tor These represent the hidden states obtained after LSTM processing of each corresponding signal;

[0025] A cross-attention mechanism is used to fuse features from each channel. The fused feature representation is as follows:

[0026]

[0027] Where Attention is the attention function, α i For attention weights, h i This represents the features extracted by LSTM from each input channel. After the fused features are passed through a fully connected layer and a softmax layer, the output is the predicted probability p for each working condition. i Where n is the total number of channels, this process can be expressed by the following formula:

[0028] p i =Softmax(W·Attention(h) vib ,h rpm ,h pres ,h tor )+b)

[0029] Where Softmax(·) is the Softmax function, W is the weight matrix, and b is the bias term. The Softmax function is defined as:

[0030]

[0031] σK represents the total number of operating condition categories. Parallel computing is used to classify and calculate multiple operating conditions; the speedup of parallel computing is expressed as:

[0032]

[0033] Among them, Time serial For serial processing time, Time parallel The parallel processing time is given by k, which represents the number of parallel processing conditions.

[0034] Further, step 3 includes:

[0035] First, basic time-domain statistics are extracted from the original vibration signal. Then, the energy distribution of the signal in the frequency domain is obtained by performing a Fourier transform on the vibration signal. The frequency domain features help to identify the main frequency components and noise interference in the signal.

[0036] The Short-Time Fourier Transform (SFT) is used to capture the spectral information of a signal as it changes over time. The SFT decomposes the signal into a series of local spectra, thus providing detailed feature information in both time and frequency dimensions. Its mathematical expression is as follows:

[0037]

[0038] Where X(τ,ω) represents the short-time Fourier transform of signal x(n) at the time-frequency point (τ,ω), x(n) represents the discrete-time signal, w(n-τ) is the window function, τ represents the time displacement parameter, ω is the angular frequency, and j is the imaginary unit.

[0039] Further, step 4 includes:

[0040] Anomaly detection is performed under different operating conditions, and parallel computing is used to accelerate the detection speed:

[0041] For vibration signals, the features of the vibration signals and the original time-series data are first extracted. Then, these two types of data are input into CNN and LSTM networks for processing, respectively. After feature extraction by each network, the two sets of features are fused using a gating attention mechanism. The fused network can automatically focus on key moments and abnormal patterns. The principle of the CNN network is as follows:

[0042] CNN f eatures = CNN(X) a1 )

[0043] Where CNN(·) represents a convolutional neural network, CNN_features represents the features output by the CNN, and X a1 These are characteristics of the input vibration signal;

[0044] The principle of LSTM is as follows:

[0045] LSTM f eatures = LSTM(X) a2 )

[0046] Where LSTM(·) represents a Long Short-Term Memory neural network, LSTM_features represents the features output by the LSTM, and X a2 It is the raw data of the input vibration signal;

[0047] The principle of gating attention mechanism is as follows:

[0048] Attention f eatures = G(CNN) f eatures, LSTM features)

[0049] Among them, Attention f Features represent the outputs of the gated attention network, where G represents the gated attention mechanism used to fuse features extracted by the CNN and LSTM networks. The fused features are then processed by a Softmax classifier to output the fault type.

[0050] p i = softmax(W·Attention) f eatures+b)

[0051] Where Softmax(·) is the Softmax function, W and b are the weights and bias terms of the classification layer of the vibration signal, and p i Output the failure probability for each category;

[0052] For pressure, speed, and torque signals, the raw data of these signals are directly input into the LSTM network for processing. The LSTM network can capture the time-dependent features of these signals, expressed as:

[0053] LSTM f eatures b =LSTM(X) b ),LSTM f eatures c =LSTM(X) c ),LSTM f eatures d =LSTM(X) d )

[0054] Among them, LSTM f eatures b This indicates the characteristics of the pressure signal output by the LSTM. f eatures c This indicates the characteristics of the speed signal output by the LSTM. f eatures d X represents the characteristics of the torque signal output by the LSTM. b Represents the rotational speed signal, X c Represents a pressure signal, X d Representing torque signals, the above signals are concatenated and then processed by a Softmax classifier to output the fault type:

[0055] p i = softmax(W′·LSTM) f eatures+b′)

[0056] Where W' and b' are the weights and biases of the classification layer that integrates pressure, speed, and torque signals, p i 'The failure probability output for each category.'

[0057] The speedup of parallel computing can be expressed by the formula:

[0058]

[0059] Among them, Time serial For serial processing time, Time parallel The parallel processing time is given by k, which represents the number of parallel processing conditions.

[0060] 1) Integrated design: to automate the entire process of data acquisition, transmission, preprocessing, feature extraction and anomaly detection.

[0061] 2) High real-time performance and high precision: The system adopts wireless data transmission and parallel processing mechanisms, combined with wavelet denoising, STFT time-frequency analysis and deep neural networks, to improve the real-time performance and accuracy of fault detection.

[0062] 3) Multimodal information fusion: Through the Cross-Attention mechanism and LSTM and CNN neural networks, the effective fusion of multi-sensor data is achieved, enhancing the ability to extract fault features.

[0063] 4) Predictive pre-shipment maintenance: This invention can detect abnormalities in the gearbox before it is installed, which can reduce the safety hazards and economic losses caused by gearbox failures later. Attached Figure Description

[0064] Figure 1 Flowchart of this invention.

[0065] Figure 2 Data filtering and multi-condition classification flowchart.

[0066] Figure 3 Flowchart of multi-condition anomaly detection network. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] Please see Figures 1-3 A method for detecting gearbox anomalies based on a break-in bench includes the following steps:

[0069] (1) Construction of the break-in test platform

[0070] The break-in bench features adjustable gearbox motor and output speeds, enabling precise control of the input speed to simulate various operating conditions (such as high-speed light load and low-speed heavy load). This capability allows the invention to adapt to diverse and complex application scenarios, providing abundant experimental data for the optimization and verification of gearbox fault detection algorithms.

[0071] For data acquisition, the platform is equipped with high-precision vibration, pressure, speed, and torque sensors for real-time monitoring of the gearbox's operating status. The pressure, speed, and torque sensors feature an integrated mounting design, employing precise fixing methods to eliminate external interference and ensure the stability and accuracy of the acquired data. Furthermore, the break-in bench was designed with experimental precision in mind, optimizing the sensor arrangement to minimize data deviations caused by installation errors, thereby improving the repeatability and reliability of the experiment.

[0072] (2) Data Acquisition and Transmission

[0073] This invention features multi-channel data acquisition capabilities, simultaneously acquiring vibration signal (a), speed signal (b), pressure signal (c), and torque signal (d), ensuring comprehensive monitoring of the gearbox's operating status. These sensors record the equipment's dynamic changes with high precision and a high sampling rate, providing a reliable basis for subsequent data analysis and anomaly detection.

[0074] The acquired multimodal sensor data is first aggregated via a CAN bus. CAN bus, as a highly reliable and interference-resistant industrial-grade communication method, can reduce signal loss while ensuring stable data transmission, thus guaranteeing data integrity and accuracy. The aggregated data is then compressed to reduce bandwidth usage and improve data transmission efficiency.

[0075] Compressed data is transmitted remotely via a 5G network. Compared to traditional wired transmission methods, 5G networks offer low latency, high bandwidth, and massive connectivity, enabling real-time transmission of large amounts of data and ensuring that remote servers can quickly receive and process experimental data.

[0076] To ensure data integrity and timeliness, the system transmits data in groups of m, with each data packet lasting n minutes. This grouped transmission method not only facilitates real-time monitoring but also enables effective data compensation and recovery in case of data loss or anomalies, ensuring the continuity and reliability of experimental data.

[0077] (3) Data Analysis

[0078] The server receives data transmitted from the break-in bench in real time and monitors the data packets at preset reception intervals to determine whether a complete set of data has been successfully received. Upon successful reception, the server automatically decompresses the data, extracting key signals such as vibration, pressure, speed, and torque from the compressed data packets. The decompressed data is automatically sorted and synchronized according to timestamps to ensure consistency of signals from different types of sensors over time, facilitating subsequent condition analysis and anomaly detection. The parsed data is stored on the server for subsequent data filtering, feature extraction, multi-condition classification, and anomaly detection.

[0079] (4) Data filtering and multi-condition classification

[0080] To improve data quality, this invention employs wavelet denoising processing on the transmitted vibration signal. Through multi-scale analysis, environmental noise and sensor errors are effectively removed, resulting in a more stable signal while preserving key characteristics of the gearbox's operating state. The denoising process for the vibration signal can be expressed by the following formula:

[0081]

[0082] in denoted as wavelet coefficients after soft thresholding, sgn(·) is the sign function representing the positive or negative sign of the input value, W(j,k) represents the k-th coefficient in the j-th level wavelet decomposition, and λ is the set threshold.

[0083] In terms of data classification, this invention employs a multimodal LSTM neural network to extract features and perform temporal modeling on data from four sensors. Assumptions: X1 is a vibration signal, X2 is a rotational speed signal, X3 is a pressure signal, and X4 is a torque signal. Features of each signal are extracted using the LSTM neural network, as expressed by the following formula:

[0084] h vib =LSTM(X1),h rpm =LSTM(X2)

[0085] h pres =LSTM(X3),h tor =LSTM(X4)

[0086] Where X1, X2, X3, and X4 represent four original time-series signals: vibration, rotational speed, pressure, and torque, respectively; LSTM(·) represents a Long Short-Term Memory network used to extract temporal features; h vib ,h rpm ,h pres ,h tor These represent the hidden states obtained after LSTM processing of each corresponding signal.

[0087] To fully utilize the complementary information from multi-sensor data, this invention employs a cross-attention mechanism to fuse features from each channel. The fused features can be expressed as:

[0088]

[0089] Where Attention is the attention function, α i h represents the attention weights, reflecting the importance of each feature in the final fusion result. i This represents the features extracted from each input channel by the LSTM. After the fused features pass through a fully connected layer and a softmax layer, the output is the predicted probability p for each working condition. i This process can be expressed by the following formula:

[0090] p i =Softmax(W·Attention(h) vib ,h rpm ,h pres ,h tor )+b)

[0091] Where Softmax(·) is the Softmax function, W is the weight matrix, and b is the bias term. The Softmax function is defined as:

[0092]

[0093] Where, σ(z) i This represents the output value of the i-th component of vector z after passing through the softmax function; z i Let z represent the i-th element in the input vector z. j Let z represent the j-th element in the input vector, K be the dimension of z (i.e., the total number of work condition categories), and e be the base of the natural logarithm.

[0094] Furthermore, to meet real-time detection requirements, this invention utilizes parallel computing to perform classification calculations for multiple operating conditions. The speedup of parallel computing can be expressed as:

[0095]

[0096] Among them, Time serial For serial processing time, Time parallel Where k represents the parallel processing time, and k represents the number of parallel processing conditions. Through this strategy, the present invention can efficiently process large-scale data in a multi-threaded environment, enabling rapid classification and identification of abnormal states under complex operating conditions.

[0097] (5) Feature extraction

[0098] First, basic time-domain statistics, such as mean, variance, skewness, kurtosis, peak factor, and kurtosis, are extracted from the original vibration signal. These characteristics reflect the overall trend and fluctuation of the signal over time, and are of great significance for the preliminary judgment of signal anomalies and trends. By performing a Fourier transform on the vibration signal, the energy distribution of the signal in the frequency domain can be obtained. Frequency domain characteristics can help identify the main frequency components and noise interference in the signal, especially abnormal vibrations in certain specific frequency bands.

[0099] To simultaneously capture the time-varying spectral information of the signal, this invention employs Short-Time Fourier Transform (STFT). STFT decomposes the signal into a series of local spectra, thereby providing detailed feature information in both time and frequency dimensions. Its mathematical expression is:

[0100]

[0101] Where X(τ,ω) represents the short-time Fourier transform of signal x(n) at the time-frequency point (τ,ω), x(n) represents the discrete-time signal, w(n-τ) is the window function, τ represents the time displacement parameter, ω is the angular frequency, and j is the imaginary unit. The time-frequency features extracted by STFT can reveal the frequency changes of the vibration signal at different times, providing more effective information for subsequent fault diagnosis.

[0102] (6) Multi-condition anomaly detection

[0103] To improve detection speed and adapt to various operating conditions, this invention employs a parallel detection scheme for multiple working conditions. Under each working condition, this invention performs specialized processing on different signals, thereby achieving efficient anomaly detection and fault diagnosis.

[0104] For vibration signals, the system first extracts the features of channel a vibration signal and the original time-series data. Then, these two types of data are input into a CNN network and an LSTM network for processing, respectively. After feature extraction by each network, a gating attention mechanism is used to fuse the two sets of features. The fused network can automatically focus on key moments and abnormal patterns. The principle of the CNN network is as follows:

[0105] CNN f eatures = CNN(X) a1 )

[0106] Where CNN(·) represents a convolutional neural network, CNN_features represents the features output by the CNN, and X a1 It is a characteristic of the input vibration signal.

[0107] The principle of LSTM is as follows:

[0108] LSTMf eatures = LSTM(X) a2 )

[0109] Where LSTM(·) represents a Long Short-Term Memory neural network, LSTM_features represents the features output by the LSTM, and X a2 It is the raw data of the input vibration signal.

[0110] The principle of gating attention mechanism is as follows:

[0111] Attention f eatures = G(CNN) f eatures, LSTM f eatures)

[0112] Among them, Attention f Features represent the outputs of the gated attention network, where G represents the gated attention mechanism used to fuse features extracted by CNN and LSTM. The fused features are then processed by a Softmax classifier to output the fault type.

[0113] p i = softmax(W·Attention) f eatures+b)

[0114] Where Softmax(·) is the Softmax function, and W and b are the weights and biases of the classification layer for the vibration signal. i Output the failure probability for each category.

[0115] For pressure, rotational speed, and torque signals, this invention directly inputs the raw data of these signals into an LSTM network for processing. LSTM networks are suitable for processing time-series data and can capture the time-dependent features of these signals.

[0116] LSTM f eatures b =LSTM(X) b ),LSTM f eatures c =LSTM(X) c ),LSTM f eatures d =LSTM(X) d )

[0117] Among them, LSTM f eatures b This indicates the characteristics of the pressure signal output by the LSTM. features c This indicates the characteristics of the speed signal output by the LSTM. f eatures d This represents the characteristics of the torque signal output by the LSTM, where X b Represents the rotational speed timing signal, X c Represents the speed and pressure signal, X d This represents the speed and torque signal. After being concatenated, the above signals are processed by a Softmax classifier to output the fault type:

[0118] p i = softmax(W′·LSTM) f eatures+b′)

[0119] Where W' and b' are the weights and biases of the classification layer that integrates pressure, speed, and torque signals, p i 'The failure probability output for each category.'

[0120] To achieve real-time detection, this invention employs a parallel computing scheme. For different operating conditions, this invention can process multiple operating condition signals simultaneously, ensuring that each condition can be analyzed promptly. This parallel detection scheme not only accelerates data processing speed but also reduces latency caused by data backlog. Through multi-threaded parallelism, computational efficiency can be significantly improved. The speedup ratio of parallel computing is expressed by the formula:

[0121]

[0122] Among them, Time serial For serial processing time, Time parallel Let k represent the parallel processing time, and k represent the number of parallel processing steps. In this way, the present invention can process more data in a shorter time, meeting the needs of real-time detection.

[0123] The gearbox will be considered normal only if the vibration test results and the pressure, speed, and torque test results are all normal.

[0124] To ensure data accuracy and detection stability, this invention also sets a sensor signal threshold. This allows the invention to monitor the sensor's status in real time; if a signal exceeds the preset threshold, the system immediately determines whether the sensor has malfunctioned and takes timely action. This approach ensures data reliability and avoids detection errors caused by sensor failure.

[0125] (7) Front-end UI display

[0126] The corresponding detection system of this invention features an intuitive and user-friendly front-end display interface that shows raw signals, signal characteristics, and fault detection results in real time, helping users quickly obtain equipment status. The interface supports multiple data display methods, including dynamic graphs and statistical charts, allowing users to intuitively understand signal changes and fault modes. Raw signals such as vibration, rotational speed, pressure, and torque signals can be displayed as graphs, and users can zoom in to view data for specific time periods. Signal characteristic display helps users identify changes in operating conditions, while detection results show fault types and probabilities, highlighting anomalies for quick problem location. Simultaneously, the interface provides real-time push notifications, promptly alerting users when anomalies occur, ensuring rapid response from engineers.

[0127] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting gearbox anomalies based on a running-in table, characterized in that, include: Step 1: Collect and transmit data, including vibration signals, rotational speed signals, pressure signals, and torque signals; Step 2: Analyze the collected data; Step 3 involves filtering and classifying the data according to various operating conditions, including: Wavelet denoising is applied to the vibration signal. Environmental noise and sensor errors are removed through multi-scale analysis. The denoising effect on the vibration signal is expressed by the following formula: in The wavelet coefficients are processed by soft thresholding, sgn(·) is the sign function, which indicates the positive or negative sign of the input value, W(j,k) represents the k-th coefficient in the j-th level wavelet decomposition, and λ is the set threshold. In terms of data classification, a multimodal LSTM neural network is used for feature extraction and time-series modeling of vibration signals, rotational speed signals, pressure signals, and torque signals. Let X1 be the vibration signal, X2 be the rotational speed signal, X3 be the pressure signal, and X4 be the torque signal. The features of each signal are extracted using the LSTM neural network, and the formula can be expressed as: Where X1, X2, X3, and X4 represent four original time-series signals: vibration, rotational speed, pressure, and torque, respectively, and LSTM(·) represents a Long Short-Term Memory network used to extract temporal features. These represent the hidden states obtained after LSTM processing of each corresponding signal; A cross-attention mechanism is used to fuse features from each channel. The fused feature representation is as follows: Where Attention is the attention function, α i For attention weights, h i This represents the features extracted by LSTM from each input channel. After the fused features are passed through a fully connected layer and a softmax layer, the output is the predicted probability p for each working condition. i Where n is the total number of channels, this process can be expressed by the following formula: Where Softmax(·) is the Softmax function, W is the weight matrix, and b is the bias term. The Softmax function is defined as: K represents the total number of operating condition categories. Parallel computing is used to classify and calculate multiple operating conditions. The speedup of parallel computing is expressed as: Among them, Time serial For serial processing time, Time parallel The parallel processing time is given by k, which represents the number of parallel processing conditions. Step 4, feature extraction of the data, including: Anomaly detection is performed under different operating conditions, and parallel computing is used to accelerate the detection speed: For vibration signals, the features of the vibration signals and the original time-series data are first extracted. Then, these two types of data are input into CNN and LSTM networks for processing, respectively. After feature extraction by each network, the two sets of features are fused using a gating attention mechanism. The fused network can automatically focus on key moments and abnormal patterns. The principle of the CNN network is as follows: Wherein, CNN(·) represents a convolutional neural network. X represents the feature output by the CNN. a1 These are characteristics of the input vibration signal; The principle of LSTM is as follows: Wherein, LSTM(·) represents a Long Short-Term Memory neural network. X represents the feature output by the LSTM. a2 It is the raw data of the input vibration signal; The principle of gating attention mechanism is as follows: in, This represents the features output from the gated attention network, where G represents the gated attention mechanism used to fuse features extracted by the CNN and LSTM networks; the fused features are then processed by a Softmax classifier to output the fault type. For pressure, speed, and torque signals, the raw data of these signals are directly input into the LSTM network for processing. The LSTM network can capture the time-dependent features of these signals, expressed as: in, This indicates the characteristics of the pressure signal output by the LSTM. This indicates the characteristics of the speed signal output by the LSTM. X represents the characteristics of the torque signal output by the LSTM. b Represents the rotational speed signal, X c Represents a pressure signal, X d Representing torque signals, the above signals are concatenated and then processed by a Softmax classifier to output the fault type: Among them W ' and b ' To integrate the weights and biases of the classification layer that incorporate pressure, speed, and torque signals, Output the failure probability for each category; The speedup of parallel computing can be expressed by the formula: Step 5: Perform multi-condition anomaly detection.

2. The gearbox anomaly detection method based on a running-in table according to claim 1, characterized in that, Step 1 includes: First, the collected data is aggregated via a CAN bus. The aggregated data is then compressed and transmitted remotely via a wireless network. During data transmission, data packets are sent in groups of m, with each packet lasting n minutes.

3. The gearbox anomaly detection method based on a running-in table according to claim 1, characterized in that, Step 2 includes: The server receives data transmitted from the break-in bench in real time and monitors the data packets at preset reception intervals to determine whether a complete set of data has been successfully received. After successfully receiving the complete data, the server automatically performs data decompression and extracts vibration signals, pressure signals, speed signals, and torque signals from the compressed data packets. The decompressed data is automatically sorted and synchronized according to timestamps to ensure that different types of sensor signals remain consistent in the time dimension.

4. The gearbox anomaly detection method based on a running-in table according to claim 1, characterized in that, Step 3 includes: First, basic time-domain statistics are extracted from the original vibration signal. Then, the energy distribution of the signal in the frequency domain is obtained by performing a Fourier transform on the vibration signal. The frequency domain features help to identify the main frequency components and noise interference in the signal. The Short-Time Fourier Transform (SFT) is used to capture the spectral information of a signal as it changes over time. The SFT decomposes the signal into a series of local spectra, thus providing detailed feature information in both time and frequency dimensions. Its mathematical expression is as follows: Where X(τ,ω) represents the short-time Fourier transform of signal x(n) at the time-frequency point (τ,ω), x(n) represents the discrete-time signal, w(n−τ) is the window function, τ represents the time displacement parameter, ω is the angular frequency, and j is the imaginary unit.

Images