A worm gear drive life test system and method

Abstract

The present application relates to the technical field of transmission component life test, in particular to a worm gear drive life test system and method, comprising: collecting drive temperature, vibration, noise signals under simulated load conditions, jointly analyzing multi-source signals in time and frequency domains and extracting multi-dimensional feature vectors. An improved deep residual network is used to construct a life prediction model, and the feature vectors are input to calculate the current health status index and the remaining useful life prediction value synchronously, and a comprehensive test report containing evaluation results and maintenance suggestions is generated according to the double indexes. The method can fully integrate multi-physical quantity degradation information, enhance deep feature learning ability, and realize accurate and comprehensive life evaluation and state monitoring.

Description

Technical Field

[0001] This invention relates to the field of transmission component life testing technology, and in particular to a worm gear drive life testing system and method. Background Technology

[0002] Traditional worm gear drive life testing often employs a single signal monitoring method, collecting only vibration or temperature data and obtaining characteristic parameters through time-domain statistics or simple frequency-domain analysis. This results in a limited data source and narrow analytical dimensions. Life prediction relies on traditional shallow models or empirical formulas, which are insufficient for characterizing long-term degradation processes.

[0003] A single signal is insufficient to cover multiple types of fault information such as temperature rise, wear, and meshing abnormalities. Temperature, vibration, and noise data are not jointly analyzed, and time-frequency domain features are independent of each other, failing to form a complete degradation characterization system. Conventional deep networks are prone to gradient vanishing, insufficient feature extraction, and weak model convergence stability during deep modeling.

[0004] Existing methods cannot achieve joint time-frequency domain feature extraction of multi-source signals, and the prediction model cannot simultaneously output the health status index and remaining service life. It is necessary to construct a multi-signal joint analysis mechanism and use an improved deep residual network to achieve simultaneous prediction of dual indicators in order to fully reflect the driver's life decay process. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and to propose a worm gear drive life testing system and method.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for testing the life of a worm gear drive, comprising: Real-time operating data of the tested worm gear drive under simulated load conditions are collected, including temperature signals, vibration signals and noise signals; The collected temperature, vibration, and noise signals are jointly analyzed in the time and frequency domains to extract multi-dimensional feature vectors. The multi-dimensional feature vectors are input into a pre-trained lifetime prediction model, which is constructed based on an improved deep residual network algorithm. The current health status index and remaining service life prediction value of the tested worm gear drive are calculated using the life prediction model. Based on the current health status index and the predicted remaining service life, a comprehensive test report containing life assessment results and maintenance recommendations is generated and output.

[0007] As a further aspect of the present invention, the acquisition of real-time operating data of the tested worm gear drive under simulated load conditions specifically includes: A programmable simulated load torque is applied to the input of the worm gear drive under test. A vibration acceleration sensor was installed on the outer ring of the worm support bearing of the worm gear drive under test to collect radial and axial vibration signals. A non-contact infrared temperature sensor was installed on the sealed housing surface of the worm gear drive under test to collect temperature signals in the key friction pair area; A sound level meter was installed at a fixed position around the worm gear drive under test to collect the noise signal during its operation. The vibration signal, temperature signal and noise signal are synchronously acquired by the data acquisition card and transmitted to the host computer.

[0008] As a further aspect of the present invention, the joint time-frequency domain analysis of the acquired temperature signal, vibration signal, and noise signal to extract multi-dimensional feature vectors specifically includes: The temperature signal, vibration signal, and noise signal are preprocessed respectively, and the preprocessing includes outlier removal, moving average filtering, and signal alignment. Short-time Fourier transform is performed on the preprocessed vibration signal and noise signal to obtain the corresponding time-frequency spectrum. From the time-frequency spectrum of the vibration signal, the energy distribution, spectral centroid, and spectral width within a preset frequency band are extracted as vibration frequency domain features. From the time-spectrum diagram of the noise signal, the sound pressure level and sound quality objective parameters of a specific octave band are extracted as noise frequency domain features; First-order difference calculation is performed on the preprocessed temperature signal to extract the temperature change rate and temperature fluctuation extreme values ​​as temperature time-domain features. The vibration frequency domain features, the noise frequency domain features, and the temperature time domain features are standardized and fused into a multi-dimensional feature vector.

[0009] As a further aspect of the present invention, the working principle of the improved deep residual network algorithm specifically includes: Construct a deep neural network backbone structure containing multiple residual blocks, each residual block containing a convolutional layer, a batch normalization layer, and an activation function layer; At the output of each residual block, a multi-scale feature extraction branch is introduced in parallel. The multi-scale feature extraction branch uses dilated convolution kernels of different sizes to perform parallel convolution operations on the input feature map of the residual block. The outputs of each branch of the multi-scale feature extraction branch are concatenated, and then the channel weights are recalibrated through the compression and attention activation module. The multi-scale fusion features after weight recalibration are added element-wise to the backbone output features of the original residual block to form an improved residual block output. At the end of the deep neural network, the high-dimensional feature map output by the last improved residual block is input into the global average pooling layer, and then connected to the fully connected layer to output the health status index of the tested worm gear drive. The channel weight recalibration after compression and attention activation module specifically includes: The feature maps input to the compression and activation attention module are subjected to global average pooling in the spatial dimension to generate a global statistical descriptor for each channel; The global statistical descriptor is input into a multilayer perceptron, which includes a dimension-reducing fully connected layer, a nonlinear activation function, and a dimension-upgrading fully connected layer. The importance weight of each channel is learned and output through the multilayer perceptron, and the number of importance weights is the same as the number of channels in the input feature map. The learned importance weights are multiplied channel by channel in the input feature map to recalibrate the channel weights of the original feature map, thereby enhancing the response of important feature channels and suppressing the response of minor feature channels.

[0010] As a further aspect of the present invention, the multi-dimensional feature vector is input into a pre-trained lifetime prediction model, which is constructed based on an improved deep residual network algorithm and specifically includes: Obtain the historical test database, which contains multiple sets of time-series multi-dimensional feature vectors of the worm gear drive from initial operation to failure, as well as the corresponding true remaining service life labels; The historical test database is divided into a training set, a validation set, and a test set. Using the improved deep residual network algorithm as the architecture, the network weight parameters of the lifetime prediction model are initialized; The life prediction model is iteratively trained using the time-series multi-dimensional feature vectors in the training set as model inputs and the corresponding real remaining lifespan labels as supervision signals. The model performance is evaluated using the validation set after each iteration. When the prediction error of the model on the validation set no longer decreases, training is stopped, and the pre-trained lifetime prediction model is saved.

[0011] As a further aspect of the present invention, the time-series multi-dimensional feature vectors in the training set are used as model input, and the corresponding true remaining useful life labels are used as supervision signals to iteratively train the life prediction model, specifically including: Extract time-series multi-dimensional feature vector samples and their corresponding true remaining lifespan labels from the training set in batches; The extracted batch samples are input into the lifetime prediction model, and forward propagation calculation is performed through the improved deep residual network algorithm to obtain the predicted remaining lifetime value of the batch. The root mean square error between the predicted remaining useful life of the batch and the corresponding actual remaining useful life label is calculated and used as the loss function value for this training iteration. An adaptive moment estimation algorithm is used to calculate the gradient of each network weight parameter in the lifetime prediction model based on the loss function value; Based on the calculated gradient, update the network weight parameters of the lifespan prediction model to complete one iteration of training. Repeat this process until the training stop condition is met.

[0012] As a further aspect of the present invention, the current health status index and remaining service life prediction value of the tested worm gear drive are calculated using the life prediction model, specifically including: The multi-dimensional feature vector corresponding to the current moment of the tested worm gear drive is input into the pre-trained life prediction model; The pre-trained lifetime prediction model performs deep nonlinear transformation and feature abstraction on the input multi-dimensional feature vector through its internal improved deep residual network. At the end of the fully connected layer of the improved deep residual network, a scalar value representing the current health status index of the tested worm gear drive is output, wherein the health status index ranges from zero to one, and a value of one represents a completely new state. Meanwhile, in another parallel fully connected layer of the improved deep residual network, the predicted remaining service life of the tested worm gear drive is output, and the predicted remaining service life is measured in time units.

[0013] As a further aspect of the present invention, after outputting the predicted remaining service life value of the tested worm gear drive, a confidence evaluation step of the predicted value is also included: Obtain the error distribution statistics of the life prediction model when predicting historical samples on the test set; Based on the error distribution statistics, the expected error range of the life prediction model under the current operating conditions is calculated. Based on the position of the current multi-dimensional feature vector of the tested worm gear actuator in the feature space, calculate its similarity with the feature distribution of historical training samples; Based on the expected error range and the similarity, the confidence level of the current output remaining useful life prediction value is calculated using a Bayesian inference method.

[0014] As a further aspect of the present invention, based on the current health status index and the predicted remaining service life, a comprehensive test report containing service life assessment results and maintenance recommendations is generated and output, specifically including: The current health status index is mapped to a qualitative health level, which includes healthy, attentive, warning, and faulty. The remaining service life prediction value is compared with the preset maintenance cycle threshold to determine whether to trigger preventive maintenance recommendations; The life assessment results, the predicted remaining lifespan, the qualitative health level, and the preventive maintenance recommendations are formatted and organized according to a preset report template. The comprehensive test report includes additional identification information, test time information, and visualization charts of key multi-dimensional feature vectors of the tested worm gear drive. The generated comprehensive test report is saved to the specified database in electronic document format, and the report content is displayed on the user interface.

[0015] As a further aspect of the present invention, the present invention also includes a worm gear drive life testing system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the worm gear drive life testing method described above.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Simultaneous time-frequency domain joint analysis of temperature, vibration and noise signals enables the fusion and characterization of multi-source monitoring data in both time and frequency dimensions. Different physical quantity signals form complementary relationships, and the extracted multi-dimensional feature vectors can fully reflect the dynamic changes and frequency distribution characteristics during the operation of the actuator. Degradation-related weak features can be effectively preserved, and the degradation information contained in the signal can be fully mined and utilized.

[0017] An improved deep residual network algorithm is used to construct a lifetime prediction model, which can maintain gradient propagation stability while deepening the network structure. It autonomously learns and maps deep degradation patterns in high-dimensional features, enhancing the model's ability to represent nonlinear, long-cycle degradation processes. By inputting multi-dimensional feature vectors into the prediction model, the current health status index and the predicted remaining lifetime can be calculated simultaneously. Real-time health status and lifetime decline trends are output simultaneously, establishing a correspondence between the two types of indicators. This enhances the completeness and consistency of the lifetime assessment results, enabling the assessment to accurately reflect the entire lifecycle degradation evolution of the driver. Attached Figure Description

[0018] Figure 1 This is a state diagram of a worm gear actuator life test method according to the present invention; Figure 2 A flowchart for collecting real-time operating data of the tested worm gear drive under simulated load conditions; Figure 3 A flowchart illustrating the work on the improved deep residual network algorithm. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1This invention provides a method for testing the lifespan of a worm gear drive. The method includes: acquiring real-time operating data of the worm gear drive under simulated load conditions, wherein the acquired real-time operating data includes temperature signals, vibration signals, and noise signals; performing time-frequency domain joint analysis on the acquired temperature signals, vibration signals, and noise signals to extract multi-dimensional feature vectors from these signals; inputting the extracted multi-dimensional feature vectors into a pre-trained lifespan prediction model, which is constructed based on an improved deep residual network algorithm; calculating the current health status index and the predicted remaining service life of the worm gear drive under test using the lifespan prediction model; and finally, generating and outputting a comprehensive test report containing lifespan assessment results and maintenance recommendations based on the calculated current health status index and the predicted remaining service life.

[0022] In one embodiment of the present invention, see [reference] Figure 2 A programmable analog load torque is applied to the input end of the worm gear drive under test. A vibration acceleration sensor is installed on the outer ring of the worm support bearing of the worm gear drive under test to collect radial and axial vibration signals. A non-contact infrared temperature sensor is installed on the sealed housing surface of the worm gear drive under test to collect temperature signals in the critical friction pair area. A sound level meter is installed at a fixed position around the worm gear drive under test to collect noise signals during its operation. The time-series data of the vibration signal, temperature signal, and noise signal are synchronously collected through a data acquisition card and transmitted to a host computer. In the feature extraction stage, the temperature signal, vibration signal, and noise signal are preprocessed respectively. The preprocessing includes outlier removal, moving average filtering, and signal alignment. Short-time Fourier transform is performed on the preprocessed vibration signal and noise signal to obtain the corresponding time-frequency spectrum. From the time-frequency spectrum of the vibration signal, the energy distribution, spectral centroid, and spectral width within a preset frequency band are extracted as vibration frequency domain features. From the time-spectrum graph of the noise signal, the sound pressure level and sound quality objective parameters at a specific octave band are extracted as noise frequency domain features. First-order difference calculation is performed on the preprocessed temperature signal to extract the temperature change rate and temperature fluctuation extreme values ​​as temperature time domain features. The vibration frequency domain features, the noise frequency domain features, and the temperature time domain features are standardized and fused into a multi-dimensional feature vector.

[0023] In practical implementation, the data acquisition and multi-dimensional feature vector extraction process for the life test of the worm gear drive under test is illustrated through an example scenario. The worm gear drive under test is mounted on a simulated load test bench. A programmable servo motor loading system is connected to the input end of the worm gear drive under test to apply a programmable simulated load torque, simulating stress conditions under actual working conditions. In this implementation, piezoelectric vibration acceleration sensors are installed at the radial and axial positions of the outer ring of the worm support bearing of the worm gear drive under test to collect radial and axial vibration signals. Furthermore, a sealing housing surface of the worm gear drive under test is positioned near the worm wheel. A non-contact infrared temperature sensor is installed in the critical friction pair area of ​​the meshing zone to collect temperature signals from the critical friction pair area. In specific implementation, a precision sound level meter is installed at a fixed position one meter away from the worm gear drive under test to collect noise signals during its operation. The timing data of the vibration signal, temperature signal, and noise signal are collected by a data acquisition card in synchronous trigger mode and transmitted to the host computer for storage at a set sampling rate. The data comparison reflects the synchronization of signals from different sensors. The sampling rate of the vibration signal is 10 kHz, the sampling rate of the temperature signal is 10 Hz, and the sampling rate of the noise signal is 44.1 kHz. The data acquisition card ensures that the timestamps of all signals are aligned.

[0024] In some embodiments, the acquired temperature, vibration, and noise signals are preprocessed. This preprocessing includes outlier removal, moving average filtering, and signal alignment. Outlier removal uses a 3-sigma criterion to identify and remove sampling points that significantly deviate from the normal range. Moving average filtering uses a moving average filter with a window length of 5 to smooth high-frequency noise in the signal. Signal alignment is based on a unified time reference provided by the data acquisition card, adjusting the time-series data of all sensor signals to the same time axis. In some embodiments, the preprocessed vibration and noise signals undergo short-time Fourier transform to obtain the corresponding time-frequency spectrogram. The short-time Fourier transform uses a Hanning window with a window length of 256 sampling points and an overlap rate of 50%, converting the one-dimensional time-series signal into a two-dimensional time-frequency matrix. Data comparison is reflected in the time-frequency spectrogram, which simultaneously displays the changes in signal frequency components over time, providing richer fault characteristic information compared to a simple time-domain waveform.

[0025] Optionally, the energy distribution, spectral centroid, and spectral width within a preset frequency band can be extracted from the time-frequency spectrum of the vibration signal as vibration frequency domain features. The preset frequency band is set according to the structural resonant frequency of the worm gear actuator under test, for example, from 0 Hz to 5000 Hz. The energy distribution is calculated as the sum of the squares of the amplitudes at each frequency point in the time-frequency spectrum within that frequency band. The spectral centroid reflects the concentration frequency of the vibration energy, and the spectral width describes the degree of dispersion of the vibration energy. Optionally, the sound pressure level and sound quality objective parameters of a specific octave band can be extracted from the time-frequency spectrum of the noise signal as noise frequency domain features. The specific octave band includes octave bands with center frequencies of 125 Hz, 250 Hz, 500 Hz, 1000 Hz, 2000 Hz, and 4000 Hz. The sound pressure level is calculated as the logarithmic amplitude of the sound pressure within each octave band. The sound quality objective parameters include roughness and sharpness, which are obtained by calculating the weighted integral of the specific frequency band in the time-frequency spectrum.

[0026] It is understandable that a first-order difference calculation is performed on the preprocessed temperature signal to extract the rate of temperature change and the extreme values ​​of temperature fluctuation as time-domain features of temperature. The rate of temperature change is obtained by dividing the difference between the temperature values ​​of adjacent sampling points by the sampling time interval. The extreme values ​​of temperature fluctuation are obtained by calculating the difference between the maximum and minimum values ​​of the temperature signal within a sliding window. It is also understandable that the vibration frequency domain features, the noise frequency domain features, and the temperature time-domain features are standardized and fused into a multi-dimensional feature vector. The standardization process uses the z-score method, subtracting the mean of historical data from each feature value and then dividing by the standard deviation to make features of different dimensions comparable. The fusion process concatenates all standardized features sequentially into a high-dimensional vector, forming a multi-dimensional feature vector. Data comparison is reflected in the fact that the fused multi-dimensional feature vector simultaneously contains time-domain, frequency-domain, and time-frequency-domain information, which, compared to single-domain features, more comprehensively characterizes the operating state of the tested worm gear drive. For example, the calculation of the spectral centroid is expressed by the following formula:

[0027] Where: symbol Indicates the centroid of the spectrum, symbol This represents the frequency value corresponding to the frequency component index k, with the symbol k. The sign represents the amplitude corresponding to the frequency component index k. This indicates the total number of frequency components.

[0028] In one embodiment of the present invention, see [reference] Figure 3A deep neural network backbone structure containing multiple residual blocks is constructed, with each residual block containing convolutional layers, batch normalization layers, and activation function layers. At the output of each residual block, a multi-scale feature extraction branch is introduced in parallel. This branch uses dilated convolutional kernels of different sizes to perform parallel convolution operations on the input feature map of the residual block. The outputs of each branch of the multi-scale feature extraction branch are concatenated, and then the channel weights are recalibrated through a compression and attention activation module. The multi-scale fused features after weight recalibration are added element-wise to the backbone output features of the original residual block to form an improved residual block output. At the end of the deep neural network, the high-dimensional feature map of the last improved residual block output is input to a global average pooling layer, then connected to a fully connected layer, outputting the health status index of the tested worm gear actuator.

[0029] The channel weight recalibration process after compression and activation of the attention module involves performing global average pooling on the feature map input to the compression and activation of the attention module in the spatial dimension to generate a global statistical descriptor for each channel. This global statistical descriptor is then input into a multilayer perceptron, which includes a dimensionality-reduced fully connected layer, a non-linear activation function, and an up-dimensional fully connected layer. The multilayer perceptron learns and outputs importance weights for each channel, with the number of importance weights being the same as the number of channels in the input feature map. The learned importance weights are then multiplied channel-by-channel with the input feature map in the channel dimension to complete the channel weight recalibration of the original feature map, enhancing the response of important feature channels and suppressing the response of less important feature channels.

[0030] In practical implementation, the improved deep residual network algorithm is used to construct the core architecture of the lifetime prediction model. Taking a network structure with four improved residual blocks as an example scenario, a deep neural network backbone structure containing multiple residual blocks is constructed. Each residual block contains a convolutional layer, a batch normalization layer, and an activation function layer. The input of the first residual block is the feature map after the multi-dimensional feature vector is transformed by an initial convolutional layer. The convolutional layer inside each residual block uses a 3x3 convolutional kernel. The batch normalization layer normalizes the convolutional output. The activation function layer uses the ReLU function to introduce nonlinearity. Data comparison shows that, compared with ordinary convolutional networks without shortened connections, residual blocks alleviate the gradient vanishing problem in deep network training through identity mapping paths, allowing the network to be built deeper to obtain stronger feature extraction capabilities.

[0031] In some embodiments, a multi-scale feature extraction branch is introduced in parallel at the output of each residual block. The multi-scale feature extraction branch uses dilated convolution kernels of different sizes to perform parallel convolution operations on the input feature map of the residual block. Taking the first improved residual block as an example, its multi-scale feature extraction branch contains three parallel dilated convolution layers with dilated convolution kernel sizes of 3x3, 5x5, and 7x5, respectively, and corresponding dilation rates of 1, 2, and 3. These parallel dilated convolution layers simultaneously convolve the input feature map of the residual block, capturing feature patterns of different scales with different receptive fields. Data comparison shows that the multi-scale feature extraction branch can simultaneously perceive local subtle features and broader contextual features, while a single-scale convolution kernel can only capture neighborhood information within a fixed range. In some embodiments, the outputs of each branch of the multi-scale feature extraction branch are concatenated, and then the channel weights are recalibrated by the compression and attention activation module. The channel concatenation operation directly connects the feature maps of the three branch outputs in the channel dimension, forming a feature map with three times the number of channels. This feature map is then input into the compression and attention activation module.

[0032] Optionally, the specific process of channel weight recalibration in the compression and activation attention module is as follows: Global average pooling is performed on the feature map input to the compression and activation attention module in the spatial dimension to generate a global statistical descriptor for each channel. The size of the input feature map is [H, W, C]. After global average pooling, a vector of dimension [1, 1, C] is generated, where each element is the average feature value of all spatial locations of the corresponding channel. Optionally, the global statistical descriptor is input to a multilayer perceptron. The multilayer perceptron contains a dimension reduction fully connected layer, a non-linear activation function, and a dimension increase fully connected layer. The dimension reduction fully connected layer compresses the C-dimensional input to C / r dimensions, where r is the compression ratio, set to 16. The non-linear activation function uses ReLU. The dimension increase fully connected layer expands the C / r-dimensional features back to C dimensions. The multilayer perceptron learns and outputs the importance weights for each channel. The number of importance weights is the same as the number of channels C in the input feature map. The learned importance weights are a C-dimensional vector, where each scalar value represents the importance of the corresponding channel.

[0033] It can be understood that the learned importance weights are multiplied channel-by-channel with the input feature map to recalibrate the channel weights of the original feature map, enhancing the response of important feature channels and suppressing the response of minor feature channels. The recalibration process involves broadcasting the C-dimensional importance weight vector to the original input feature map along the channel dimension, that is, multiplying all spatial position values ​​of each channel of the feature map by the weight scalar corresponding to that channel. It can also be understood that the multi-scale fused features after weight recalibration are added element-by-elementally with the backbone output features of the original residual block to form the improved residual block output. The backbone output features of the original residual block refer to the features obtained through weight recalibration. The feature map obtained after passing through the standard convolutional layer, batch normalization layer, and activation function layer within the residual block is element-wise summed to add the values ​​of the multi-scale fused features and the backbone output features at corresponding positions, forming a feature map that incorporates multi-scale attention information. This serves as the final output of the improved residual block. At the end of the deep neural network, the high-dimensional feature map output from the last improved residual block is input to a global average pooling layer, followed by a fully connected layer, outputting the health status index of the tested worm gear actuator. The global average pooling layer compresses the spatial dimension of the feature map to 1x1, and the fully connected layer maps the compressed feature vector to the final scalar output. The calculation of channel weight recalibration can be expressed by the following formula:

[0034] Where: symbol This represents the learned importance weight scalar value for the c-th channel, with the sign... This represents the Sigmoid activation function, symbol... This represents the weight matrix of the upgraded fully connected layer, denoted by the symbol. Represents the ReLU activation function, symbol This represents the weight matrix of the dimensionality-reduced fully connected layer, denoted by [symbol]. This represents the global statistical descriptor for the c-th channel obtained through global average pooling.

[0035] In one embodiment of the present invention, a historical test database is obtained, which contains multiple sets of time-series multi-dimensional feature vectors of the entire process of worm gear drives from initial operation to failure, and corresponding true remaining service life labels. The historical test database is divided into a training set, a validation set, and a test set. Using the improved deep residual network algorithm as the architecture, the network weight parameters of the life prediction model are initialized. The time-series multi-dimensional feature vectors in the training set are used as model input, and the corresponding true remaining service life labels are used as supervision signals to iteratively train the life prediction model. After each iteration, the model performance is evaluated using the validation set. When the prediction error of the model on the validation set no longer decreases, training is stopped, and the pre-trained life prediction model is saved. The specific process of model iterative training is as follows: time-series multi-dimensional feature vector samples and their corresponding true remaining service life labels are extracted from the training set in batches. The extracted batch samples are input into the life prediction model, and forward propagation calculation is performed through the improved deep residual network algorithm to obtain the predicted remaining service life value of the batch. The root mean square error (RMSE) between the predicted remaining useful life (UPS) value of the batch and the corresponding actual UPS label is calculated and used as the loss function value for this training iteration. An adaptive moment estimation algorithm is used to calculate the gradient of each network weight parameter in the UPS prediction model based on the loss function value. The network weight parameters of the UPS prediction model are updated according to the calculated gradients, completing one iteration of training. This process is repeated until the training stop condition is met.

[0036] In practical implementation, the construction and training process of the life prediction model begins with the preparation of a historical test database. An example scenario involves a historical test database containing full life cycle test data of fifty worm gear drives of the same model. The historical test database contains multiple sets of time-series multi-dimensional feature vectors of the worm gear drives from initial operation to failure, along with corresponding true remaining service life labels. Data records for each drive start from a brand-new state, running under various simulated load conditions until functional failure occurs. During operation, multi-dimensional feature vectors are collected and extracted at fixed time intervals. Each feature vector corresponds to a time point. The true remaining service life label is calculated based on the total running time and the current time, and is represented as the number of time units remaining until failure. See Table 1 for the data structure of the historical database. Table 1: Historical Test Database Partition Table

[0037] In some embodiments, the historical test database is divided into a training set, a validation set, and a test set. The division follows the principle of stratified sampling to ensure that the training set, validation set, and test set contain data under different load conditions and different failure modes, so as to avoid the model overfitting to specific conditions. The training set is used to learn the model weight parameters, the validation set is used to monitor the model performance during training and determine when to stop training, and the test set is used to finally evaluate the generalization ability of the trained lifetime prediction model. In some embodiments, an improved deep residual network algorithm is used as the architecture to initialize the network weight parameters of the lifetime prediction model. The network architecture adopts a deep structure containing four improved residual blocks. Each improved residual block is configured as described in Embodiment 2. The weight parameters are randomly initialized using the He normal distribution method, and the bias parameters are initialized to zero.

[0038] Optionally, the time-series multi-dimensional feature vectors from the training set are used as model input, and the corresponding real remaining lifespan labels are used as supervision signals to iteratively train the lifespan prediction model. After each iteration, the model performance is evaluated using a validation set. The total number of iterations is set to 200. Each training round uses all samples in the training set once. After each training round, the prediction error is calculated using the data from the validation set to monitor the model's performance on the unseen validation set. Optionally, when the model's prediction error on the validation set no longer decreases, training is stopped, and the pre-trained lifespan prediction model is saved. The condition for stopping training is set to the point that the loss function value of the validation set no longer decreases within 20 consecutive training rounds. At this point, the training process is terminated early, and the model weight parameters corresponding to the lowest loss on the validation set are saved as the final trained lifespan prediction model.

[0039] The iterative training process of the model can be understood as follows: Multi-dimensional feature vector samples and their corresponding true remaining useful lives labels are extracted from the training set in batches, with a batch size of 32. Before each training round, the order of the training set samples is randomly shuffled to ensure the randomness of batch sampling. The extracted batch samples are then input into the lifetime prediction model, and forward propagation is performed using an improved deep residual network algorithm to obtain the predicted remaining useful lives for each batch. During forward propagation, the 32 multi-dimensional feature vector samples from each batch are simultaneously input into the network. After passing through multiple improved residual blocks and nonlinear transformations, 32 predicted remaining useful lives are finally output at the fully connected layer. The root mean square error (RMSE) between the predicted remaining useful lives of each batch and the corresponding true remaining useful lives label is calculated and used as the loss function value for this training iteration. The formula for calculating the RMSE is:

[0040] Where: symbol Represents the loss function value, symbol Indicates batch size, symbol This represents the model's predicted remaining useful life for the i-th sample in the batch, with the symbol... This represents the true remaining useful life label corresponding to the i-th sample in the batch. An adaptive moment estimation algorithm (IME) is used to calculate the gradients of the network weight parameters in the lifetime prediction model based on the loss function value. The initial learning rate of the IME is set to 0.001, and the decay rates for the first and second moments are set to 0.9 and 0.999, respectively. Based on the calculated gradients, the network weight parameters of the lifetime prediction model are updated, completing one iteration of training. This process is repeated until the stopping condition is met. The weight parameter updates employ the gradient descent principle, combined with the adaptive learning rate calculated by the IME, to iteratively adjust the weights and bias parameters in the network to minimize the loss function value.

[0041] In one embodiment of the present invention, the multi-dimensional feature vector corresponding to the current moment of the worm gear drive under test is input into the pre-trained life prediction model. The pre-trained life prediction model performs deep nonlinear transformation and feature abstraction on the input multi-dimensional feature vector through its internal improved deep residual network. At the end of the fully connected layer of the improved deep residual network, a scalar value representing the current health status index of the worm gear drive under test is output, where the health status index ranges from zero to one, with a value of one representing a completely new state. Simultaneously, at another parallel fully connected layer of the improved deep residual network, the remaining service life prediction value of the worm gear drive under test is output, and the remaining service life prediction value is measured in time units. After outputting the remaining service life prediction value, a confidence evaluation step for the prediction value is performed. The error distribution statistics of the life prediction model when predicting historical samples on the test set are obtained. Based on the error distribution statistics, the expected error range of the prediction result of the life prediction model under the current operating conditions is calculated. By combining the position of the current multi-dimensional feature vector of the tested worm gear actuator in the feature space, its similarity to the feature distribution of historical training samples is calculated. Based on the expected error range and the similarity, the confidence level of the current output remaining service life prediction value is calculated using a Bayesian inference method.

[0042] In practical implementation, the current health status index and remaining service life prediction value of the tested worm gear drive are calculated through a life prediction model. Taking a specific test scenario as an example, assuming the device number of the tested worm gear drive is WT-2023-005, the multi-dimensional feature vector corresponding to the current moment of the tested worm gear drive is input into the pre-trained life prediction model. This feature vector is extracted after processing the signals collected by the device WT-2023-005 under a specific simulated load. This feature vector is a one-dimensional array containing vibration frequency domain features, noise frequency domain features, and temperature time domain features. The data comparison shows that the input feature vector reflects the comprehensive state of the device WT-2023-005 at the current specific operating moment, while the model training uses the time-series feature vectors of many devices in the historical database throughout their entire life cycle.

[0043] In some embodiments, the pre-trained lifetime prediction model uses its internal improved deep residual network to perform deep nonlinear transformations and feature abstraction on the input multidimensional feature vector. The input multidimensional feature vector is first reshaped into a two-dimensional structure suitable for convolutional network processing, and then sequentially passed through multiple improved residual blocks. Each improved residual block performs convolution, batch normalization, activation, and multi-scale feature fusion and channel attention recalibration operations, gradually mapping the original features to high-level abstract features. Data comparison shows that the improved deep residual network's deep nonlinear transformation capability enables the model to transform the original multidimensional feature vector... The system automatically learns complex patterns related to lifespan degradation, which simple linear models or shallow networks struggle to capture. In some embodiments, at the end of the fully connected layer of the improved deep residual network, a scalar value representing the current health status index of the tested worm gear drive is output, where the health status index ranges from zero to one, with one representing a completely new state. Simultaneously, at another parallel fully connected layer of the improved deep residual network, a predicted remaining lifespan of the tested worm gear drive is output, measured in time units, for example, a predicted remaining lifespan of 350 hours.

[0044] Optionally, after outputting the predicted remaining service life of the tested worm gear drive, a confidence evaluation step is also included. This step obtains the error distribution statistics of the life prediction model when predicting historical samples on the test set. The test set contains a set of historical drive life data that were not involved in model training. The absolute error between the model's predicted remaining service life values ​​and the actual remaining service life values ​​for these data is recorded. The error distribution statistics are shown in Table 2, which illustrates the sample size distribution within different absolute error intervals. Table 2: Error Distribution Statistics of Lifetime Prediction Model on Test Set

[0045] Optionally, based on error distribution statistics, the expected error range of the life prediction model under the current operating conditions is calculated. This expected error range can be set to cover a certain proportion of historical prediction errors. For example, based on the table above, it can be calculated that 80% of historical prediction samples have an absolute error within 50 hours. Therefore, under similar current operating conditions, the expected error range can be set to [-50 hours, +50 hours]. It can be understood that, combined with the position of the current multi-dimensional feature vector of the tested worm gear drive in the feature space, its similarity to the feature distribution of historical training samples is calculated. The similarity calculation can use an exponential form based on Mahalanobis distance. First, the Mahalanobis distance from the current multi-dimensional feature vector to the center of the historical training set feature distribution is calculated. Then, this distance is mapped to a similarity metric between 0 and 1 through an exponential decay function. It can be understood that, based on the expected error range and the similarity, the confidence level of the current output remaining life prediction value is calculated using Bayesian inference methods. One way to express the confidence level calculation is as follows:

[0046] Where: symbol This represents the calculated confidence level, with a value between 0 and 1, and the sign... This represents the similarity metric between the current multi-dimensional feature vector and the feature distribution of historical training samples. Its value ranges from 0 to 1, and the sign is... This represents half the width of the expected error interval calculated based on the error distribution statistics, i.e., the error radius, denoted by [symbol missing]. This represents a reference error value used to determine the error radius. Normalize to make To become a dimensionless quantity, with reference error value The average of historical prediction errors can be used. This formula reflects the logical relationship that the higher the similarity and the smaller the normalized relative error radius, the higher the calculated confidence level. Furthermore, both sides of the formula are dimensionless values ​​with consistent dimensions.

[0047] In one embodiment of the present invention, the current health status index is mapped to a qualitative health level, which includes healthy, alert, warning, and fault. The predicted remaining service life is compared with a preset maintenance cycle threshold to determine whether preventive maintenance recommendations are triggered. The service life assessment results, the predicted remaining service life, the qualitative health level, and the preventive maintenance recommendations are formatted and organized according to a preset report template. The comprehensive test report includes the identification information of the tested worm gear drive, test time information, and visualization charts of key multi-dimensional feature vectors. The generated comprehensive test report is saved to a designated database in electronic document format, and the report content is displayed on the user interface.

[0048] In practical implementation, the process of generating and outputting a comprehensive test report based on the current health status index and the predicted remaining service life is described using a specific example. Assume the identification information of the tested worm gear drive is WT-2023-005, the model calculates a current health status index of 0.72, and the predicted remaining service life is 280 hours. In practice, the current health status index is mapped to a qualitative health level, which includes healthy, alert, warning, and fault. The mapping process is completed through a preset threshold range. The health status index is defined as follows: the interval [0.85, 1.00] corresponds to the "healthy" level, the interval [0.60, 0.85) corresponds to the "attention" level, the interval [0.30, 0.60) corresponds to the "warning" level, and the interval [0.00, 0.30) corresponds to the "fault" level. For the health status index of 0.72, it falls within the interval [0.60, 0.85), and is therefore mapped to the "attention" level. The data comparison shows that the qualitative "attention" level provides users with a more intuitive understanding of the status than the numerical value of 0.72.

[0049] In some embodiments, the remaining useful life prediction value is compared with a preset maintenance cycle threshold to determine whether to trigger a preventive maintenance recommendation. The preset maintenance cycle threshold is set according to the equipment maintenance procedure. For example, the trigger threshold for preventive maintenance is specified as the remaining useful life prediction value being less than 300 hours. When the remaining useful life prediction value (280 hours) is lower than the preset maintenance cycle threshold (300 hours), a preventive maintenance recommendation is triggered, and the recommendation content is "It is recommended to arrange preventive maintenance in the near future". In some embodiments, the life assessment results, remaining useful life prediction value, qualitative health level and preventive maintenance recommendation are formatted and organized according to a preset report template. The preset report template is a structured electronic document framework that includes a title, equipment information, test summary, detailed results and recommendations. The formatting and organization process is to automatically fill the various result data obtained from calculation and judgment into the corresponding fields of the report template.

[0050] Optionally, the comprehensive test report may include additional identification information for the tested worm gear drive, test time information, and visualization charts of key multi-dimensional feature vectors. Identification information includes the equipment model WT-2023-005, the test time information is the report's generation timestamp, and the visualization charts of key multi-dimensional feature vectors are presented as line graphs or radar charts, showing the trends of major characteristic values ​​over a recent time period, such as the vibration spectrum centroid, total sound pressure level, and bearing temperature. Optionally, the generated comprehensive test report may be saved as an electronic document to a specified database, and its content displayed on the user interface. The electronic document format is PDF, and the save path is a specified server folder. The database records the report's file path and metadata. The user interface is a graphical application that loads and renders the PDF report content in the report display area of ​​the main window for user viewing. The report generation and saving process can be expressed logically as follows:

[0051] Where: symbol This refers to the final generated integrated test report document, represented by the symbol... This represents a process function that formats the data according to a preset template. (Symbol) Indicates a preset report template, symbol This indicates the identification information of the worm gear drive being tested, represented by the symbol. Indicates test time information, symbol This represents a summary of life assessment results, indicated by the symbol. Represents the predicted remaining useful life, symbol Indicates qualitative health level, symbol Indicates a preventative maintenance recommendation, symbol A visual chart representing key multi-dimensional feature vectors. This can be understood as a formatting process function. Systematically integrate all input elements within the parentheses into the report template. The corresponding location will ultimately output a complete report document. .

[0052] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for testing the lifespan of a worm gear drive, characterized in that, Includes the following steps: Real-time operating data of the tested worm gear drive under simulated load conditions is collected, and the real-time operating data includes temperature signals, vibration signals and noise signals; The collected temperature, vibration, and noise signals are jointly analyzed in the time and frequency domains to extract multi-dimensional feature vectors. The multi-dimensional feature vectors are input into a pre-trained lifetime prediction model, which is constructed based on an improved deep residual network algorithm. The current health status index and remaining service life prediction value of the tested worm gear drive are calculated using the life prediction model. Based on the current health status index and the predicted remaining service life, a comprehensive test report containing life assessment results and maintenance recommendations is generated and output.

2. The method for testing the life of a worm gear drive according to claim 1, characterized in that, The acquisition of real-time operating data of the tested worm gear drive under simulated load conditions specifically includes: A programmable simulated load torque is applied to the input of the worm gear drive under test. A vibration acceleration sensor was installed on the outer ring of the worm support bearing of the worm gear drive under test to collect radial and axial vibration signals. A non-contact infrared temperature sensor was installed on the sealed housing surface of the worm gear drive under test to collect temperature signals in the key friction pair area; A sound level meter was installed at a fixed position around the worm gear drive under test to collect the noise signal during its operation. The vibration signal, temperature signal and noise signal are synchronously acquired by the data acquisition card and transmitted to the host computer.

3. The method for testing the life of a worm gear drive according to claim 1, characterized in that, The process of performing joint time-frequency domain analysis on the collected temperature, vibration, and noise signals to extract multi-dimensional feature vectors specifically includes: The temperature signal, vibration signal, and noise signal are preprocessed respectively, and the preprocessing includes outlier removal, moving average filtering, and signal alignment. Short-time Fourier transform is performed on the preprocessed vibration signal and noise signal to obtain the corresponding time-frequency spectrum. From the time-frequency spectrum of the vibration signal, the energy distribution, spectral centroid, and spectral width within a preset frequency band are extracted as vibration frequency domain features. From the time-spectrum diagram of the noise signal, the sound pressure level and sound quality objective parameters of a specific octave band are extracted as noise frequency domain features; First-order difference calculation is performed on the preprocessed temperature signal to extract the temperature change rate and temperature fluctuation extreme values ​​as temperature time-domain features. The vibration frequency domain features, the noise frequency domain features, and the temperature time domain features are standardized and fused into a multi-dimensional feature vector.

4. The method for testing the life of a worm gear drive according to claim 1, characterized in that, The working principle of the improved deep residual network algorithm specifically includes: Construct a deep neural network backbone structure containing multiple residual blocks, each residual block containing a convolutional layer, a batch normalization layer, and an activation function layer; At the output of each residual block, a multi-scale feature extraction branch is introduced in parallel. The multi-scale feature extraction branch uses dilated convolution kernels of different sizes to perform parallel convolution operations on the input feature map of the residual block. The outputs of each branch of the multi-scale feature extraction branch are concatenated, and then the channel weights are recalibrated through the compression and attention activation module. The multi-scale fusion features after weight recalibration are added element-wise to the backbone output features of the original residual block to form an improved residual block output. At the end of the deep neural network, the high-dimensional feature map output by the last improved residual block is input into the global average pooling layer, and then connected to the fully connected layer to output the health status index of the tested worm gear drive. The channel weight recalibration after compression and attention activation module specifically includes: The feature maps input to the compression and activation attention module are subjected to global average pooling in the spatial dimension to generate a global statistical descriptor for each channel; The global statistical descriptor is input into a multilayer perceptron, which includes a dimension-reducing fully connected layer, a nonlinear activation function, and a dimension-upgrading fully connected layer. The importance weight of each channel is learned and output through the multilayer perceptron, and the number of importance weights is the same as the number of channels in the input feature map. The learned importance weights are multiplied channel by channel in the input feature map to recalibrate the channel weights of the original feature map, thereby enhancing the response of important feature channels and suppressing the response of minor feature channels.

5. The method for testing the life of a worm gear drive according to claim 1, characterized in that, The multi-dimensional feature vectors are input into a pre-trained lifetime prediction model, which is constructed based on an improved deep residual network algorithm and specifically includes: Obtain the historical test database, which contains multiple sets of time-series multi-dimensional feature vectors of the worm gear drive from initial operation to failure, as well as the corresponding true remaining service life labels; The historical test database is divided into a training set, a validation set, and a test set. Using the improved deep residual network algorithm as the architecture, the network weight parameters of the lifetime prediction model are initialized; The life prediction model is iteratively trained using the time-series multi-dimensional feature vectors in the training set as model inputs and the corresponding real remaining lifespan labels as supervision signals. The model performance is evaluated using the validation set after each iteration. When the prediction error of the model on the validation set no longer decreases, training is stopped, and the pre-trained lifetime prediction model is saved.

6. The method for testing the life of a worm gear drive according to claim 5, characterized in that, Using the time-series multi-dimensional feature vectors from the training set as model input and the corresponding true remaining useful life labels as supervision signals, the life prediction model is iteratively trained, specifically including: Extract time-series multi-dimensional feature vector samples and their corresponding true remaining lifespan labels from the training set in batches; The extracted batch samples are input into the lifetime prediction model, and forward propagation calculation is performed through the improved deep residual network algorithm to obtain the predicted remaining lifetime value of the batch. The root mean square error between the predicted remaining useful life of the batch and the corresponding actual remaining useful life label is calculated and used as the loss function value for this training iteration. An adaptive moment estimation algorithm is used to calculate the gradient of each network weight parameter in the lifetime prediction model based on the loss function value; Based on the calculated gradient, update the network weight parameters of the lifespan prediction model to complete one iteration of training. Repeat this process until the training stop condition is met.

7. The method for testing the life of a worm gear drive according to claim 1, characterized in that, The current health status index and predicted remaining service life of the tested worm gear drive are calculated using the life prediction model, specifically including: The multi-dimensional feature vector corresponding to the current moment of the tested worm gear drive is input into the pre-trained life prediction model; The pre-trained lifetime prediction model performs deep nonlinear transformation and feature abstraction on the input multi-dimensional feature vector through its internal improved deep residual network. At the end of the fully connected layer of the improved deep residual network, a scalar value representing the current health status index of the tested worm gear drive is output, wherein the health status index ranges from zero to one, and a value of one represents a completely new state. Meanwhile, in another parallel fully connected layer of the improved deep residual network, the predicted remaining service life of the tested worm gear drive is output, and the predicted remaining service life is measured in time units.

8. The method for testing the life of a worm gear drive according to claim 7, characterized in that, After outputting the predicted remaining service life of the tested worm gear drive, the method also includes a confidence assessment step for the predicted value: Obtain the error distribution statistics of the life prediction model when predicting historical samples on the test set; Based on the error distribution statistics, the expected error range of the life prediction model under the current operating conditions is calculated. Based on the position of the current multi-dimensional feature vector of the tested worm gear actuator in the feature space, calculate its similarity with the feature distribution of historical training samples; Based on the expected error range and the similarity, the confidence level of the current output remaining useful life prediction value is calculated using a Bayesian inference method.

9. The method for testing the life of a worm gear drive according to claim 1, characterized in that, Based on the current health status index and the predicted remaining service life, a comprehensive test report is generated and output, including service life assessment results and maintenance recommendations, specifically including: The current health status index is mapped to a qualitative health level, which includes healthy, attentive, warning, and faulty. The remaining service life prediction value is compared with the preset maintenance cycle threshold to determine whether to trigger preventive maintenance recommendations; The life assessment results, the predicted remaining lifespan, the qualitative health level, and the preventive maintenance recommendations are formatted and organized according to a preset report template. The comprehensive test report includes additional identification information, test time information, and visualization charts of key multi-dimensional feature vectors of the tested worm gear drive. The generated comprehensive test report is saved to the specified database in electronic document format, and the report content is displayed on the user interface.

10. A worm gear drive life testing system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the worm gear drive life test method according to any one of claims 1 to 9.

Images