Reduction gearbox input shaft end-to-end speed adaptive inference method based on variable length sequence

Abstract

The present application belongs to the technical field of rotating machinery state monitoring and industrial deployment, and particularly relates to a reducer input shaft end-to-end rotating speed adaptive inference method based on variable-length sequence. In view of the fact that actual monitoring data is a continuous sequence of any length, and a deep learning model usually relies on fixed-length input, and traditional sliding window inference is prone to introducing phase discontinuity and splicing artifacts at the boundary, the present application establishes a deployment paradigm of "fixed-length training and variable-length inference". The continuous variable-length sequence is divided by a sliding window with an overlap rate, and a backtracking interception strategy is designed for the case of insufficient end sequence. After batch inference using a pre-trained model, the local prediction results are integrated and smoothed by using an overlap addition mechanism and a weighted average algorithm. The present application effectively eliminates the Gibbs effect at the window boundary, ensures the phase integrity of the terminal, and realizes seamless and smooth rotating speed curve reconstruction of continuous monitoring data of any length.

Description

Technical Field

[0001] This invention belongs to the field of rotating machinery condition monitoring and industrial deployment technology, specifically relating to an end-to-end adaptive inference method for the input shaft speed of a gearbox based on a variable-length sequence. Background Technology

[0002] In modern industrial IoT and edge computing scenarios, extremely high real-time and continuous requirements are placed on the health monitoring of large and complex rotating electromechanical equipment such as port cranes and wind turbines. These heavy machines typically operate under non-stationary conditions such as frequent start-stop cycles, variable loads, and speed changes during actual service. The non-stationary vibration signals of core transmission components such as gearboxes contain rich state information, and accurately acquiring the transient speed of the equipment is a prerequisite for order tracking, feature decoupling, and subsequent fault diagnosis.

[0003] In recent years, end-to-end deep learning sequence models (such as convolutional neural networks, recurrent neural networks, and their variants) have demonstrated outstanding performance in rotational speed estimation and signal representation. These methods can adaptively extract mapping relationships directly from raw vibration signals, eliminating reliance on traditional tachometer hardware. However, in practical industrial edge deployments, these models face a severe technical bottleneck of "input size mismatch." Deep learning models are limited by the batch processing mechanism during the training phase and the design of the receptive field or memory state of the network structure, and their input is usually strictly fixed to short segments of a specific length; while the data collected by actual sensors at the monitoring site, after effective segment selection, is usually a continuous, non-periodic, and arbitrarily long sequence of unknown length.

[0004] To address this issue, traditional engineering processing methods typically employ a simple "sliding window" strategy for inference, which involves directly cutting long signals into multiple fixed-length short segments, feeding each segment into the model for processing, and then concatenating the outputs. However, this hard-concatenation method inevitably introduces phase discontinuities and concatenation artifacts (Gibbs effect) at the window boundaries, severely disrupting the smoothness of the rotational speed estimation curve and the continuity of the physical time sequence, thereby affecting the accuracy of subsequent order tracking and state monitoring. Therefore, there is an urgent need for an adaptive inference reconstruction algorithm that can break the strict limitations of deep learning models on input size and achieve seamless concatenation of variable-length data. Summary of the Invention

[0005] To address the limitations of existing technologies, this invention proposes an end-to-end adaptive inference method for the input shaft speed of a gearbox based on variable-length sequences. This method establishes an engineering deployment paradigm of "fixed-length training and variable-length inference," achieving seamless speed estimation of continuously monitored data of arbitrary lengths through a sliding window segmentation with overlap rate combined with a backtracking filling strategy, and an overlap summation mechanism based on weighted averages.

[0006] The technical solution of this invention is as follows: An adaptive inference method for end-to-end speed of a gearbox input shaft based on a variable-length sequence includes the following steps: Step 1: During the offline model training phase, long-term single-channel vibration signals and their aligned true rotational speed curves of the gearbox input shaft are acquired. Specifically, using an accelerometer and a photoelectric speed sensor installed near the gearbox input shaft bearing housing (or housing), historical operating data of the electromechanical equipment under various non-stationary operating conditions are simultaneously acquired. The acquired discretized acceleration data is converted into a one-dimensional long-term single-channel vibration signal. Simultaneously, by calculating the high-frequency pulse signal recorded by the photoelectric speed sensor, a high-precision transient actual rotational speed curve of the equipment is obtained, which is used as the true label for supervised learning.

[0007] Step 2: Using a fixed window strategy, the long-time single-channel vibration signal obtained in Step 1 and its strictly aligned true rotational speed curve are synchronously divided into time-divided segments. Fixed-length sample pairs were used. The segmented vibration segments underwent preprocessing such as mean removal and Z-score standardization to finally divide the data into training and validation sets, thus constructing data that meets the requirements for batch training of deep learning networks.

[0008] Step 3: Based on the data from Step 2, construct an end-to-end deep network model for the input shaft of the gearbox, oriented towards one-dimensional non-stationary time-series signals. Addressing the strong non-stationarity and complex modulation characteristics of the gearbox vibration signals under frequent load and speed variations, a deep architecture with strong long-sequence modeling capabilities, such as a bidirectional state-space model or a long short-term memory network, is employed. Due to the inherent network architecture and internal parameter mechanism, the input time dimension of the model is strictly fixed to a length. Furthermore, a single forward propagation can only map the output corresponding to a length of [length missing]. The local rotational speed prediction sequence. This model, with its rigid constraint on fixed input dimensions, constitutes the prerequisite for subsequent block-based adaptive inference and reconstruction of arbitrarily variable-length continuous sequences.

[0009] Step 4: In the online inference phase, obtain the variable-length continuous vibration tensor to be monitored. Define a sliding window mechanism with overlap ratio to transfer the original variable-length continuous vibration tensor... This is decomposed into a series of fixed-length observation subsequences suitable for the end-to-end depth network model input of the gearbox input shaft, as described in step 3. The time length of the sliding window is assumed to be consistent with the model input constraints, i.e., ... The sliding step size is And set the overlap rate between adjacent windows. Then the step size satisfies By performing overlapping sliding truncation on the complete input tensor along the time dimension, the resulting [number] [unit]... Local fixed-length observation subsequence Defined as: Step 5: Determine if the sliding window has reached the physical end of the monitored data sequence. If the sliding window reaches the end and the remaining time step length is insufficient for the fixed input length required by the model, this step is considered complete. To avoid introducing non-physical high-frequency noise during conventional zero-padding, a backtracking truncation strategy is adopted: forcing the original sequence to use its physical end time point. Starting from the endpoint, extract a complete segment of length in reverse order. The data segment is used as the last subsequence, i.e., the tail subsequence, in this batch decomposition. Formal definition: in, This is the maximum index value of the local subsequence generated after sequence decomposition.

[0010] Step 6: Combine all the subsequences generated in step 4 with the tail subsequence generated in step 5 through compensation. The sequences are merged to form a complete set of subsequences to be inferred. This set is then batch-propagated into the end-to-end speed estimation model of the gearbox input shaft described in step 3 to obtain a set of local speed estimation sequences corresponding to each local observation window. .

[0011] Step 7: Because Step 4 employed a sliding window mechanism with overlap, most sampling points on the time axis simultaneously exist in multiple adjacent local rotational speed estimation sequences. To eliminate splicing artifacts at window boundaries when the model independently processes each local sequence, a weighted averaging mechanism based on coverage counts is used to integrate and smooth the local prediction results. The formal description of the global adaptive reconstruction process is as follows: in, For global refactoring at a specific point in time The final transient speed prediction value; For the first Each local window corresponds to the output at a specific time point. The predicted local rotational speed; As an indicator function, when time point In the first When within the coverage area of ​​a local window, Otherwise, it is 0.

[0012] Step 8: Output the reconstructed global smooth speed curve This continuous speed curve is used to perform order tracking analysis and condition monitoring of non-stationary vibration signals of industrial machinery, enabling seamless speed estimation of continuous monitoring data of arbitrary length.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention breaks away from the strict limitations of traditional deep learning models on fixed input sizes, proposing an adaptive inference reconstruction algorithm based on overlapping sliding windows. Through overlapping addition and weighted averaging mechanisms, it completely eliminates phase abrupt changes and artifacts at window boundaries caused by traditional stitching methods, ensuring the smoothness of the global rotational speed curve. Simultaneously, the unique backtracking truncation strategy perfectly solves the problem of filler noise due to insufficient sequence length, ensuring the phase integrity of the physical waveform terminal. This method significantly improves the algorithm's adaptability to non-standard duration monitoring data in real-world industrial scenarios, providing an ideal engineering deployment solution for real-time, seamless status monitoring at the edge of complex equipment. Attached Figure Description

[0014] Figure 1 This is an overall adaptive reasoning flowchart of an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the power transmission chain structure of the quay crane lifting mechanism applied in an embodiment of the present invention. In the figure, 1 is a cable reel; 2 is a drive motor; and 3 is a parallel shaft gearbox.

[0016] Figure 3 This is an example of the vibration signal time-domain waveform diagram in an embodiment of the present invention.

[0017] Figure 4 This is a comparison chart of the actual rotational speed and the predicted rotational speed in an embodiment of the present invention.

[0018] Figure 5 This is the result of the order spectrum analysis verification obtained from the calculation of the actual rotational speed and the predicted rotational speed in the embodiment of the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0020] This embodiment uses the actual online monitoring task of the hoisting mechanism of an in-service quay crane at a large container port as an example for verification. The actual operating conditions of the quay crane hoisting transmission system are extremely complex, frequently experiencing alternating operation cycles such as full-load hoisting and empty hook lowering, accompanied by strong non-steady impact loads and complex meshing frequency interference inside the gearbox. In order to achieve effective state monitoring of this transmission system, it is necessary to segment and extract the segments in which the equipment is in an effective operating state from the original continuously acquired signals, so that the vibration acceleration data to be analyzed ultimately presents as a series of variable-length sequences of varying lengths.

[0021] The power transmission chain of the hoisting mechanism, such as Figure 2 As shown, the hoisting transmission system of this quay crane faces severe non-steady impact loads and strong background noise interference, and operates for a long time in typical loading and unloading operation cycles such as full-load hoisting, empty hook lowering, hovering positioning, and frequent acceleration and deceleration. The power transmission chain system of this hoisting mechanism is driven by dual motors, and the power is transmitted to the drum through a four-stage parallel shaft gearbox, which in turn drives the spreader to complete the vertical lifting operation. Photoelectric speed sensors and vibration acceleration sensors are deployed at key parts of the transmission chain. The data acquisition system synchronously records vibration signals and speed pulses at a sampling frequency of 12.8 kHz, constructing a measured variable length dataset covering the entire operation process.

[0022] To verify the proposed end-to-end adaptive inference method for the input shaft speed of the gearbox based on a variable-length sequence, as follows: Figure 1 As shown, the specific implementation and testing steps are as follows: Step 1: Using vibration acceleration sensors and photoelectric speed sensors deployed on the hoisting drive chain of the quay crane, historical operating data of the equipment under various typical non-stationary working conditions, such as full-load hoisting, empty hook lowering, and hovering, is collected synchronously. First, by processing the high-frequency pulse signals recorded by the photoelectric sensors, a high-precision actual transient speed curve of the equipment is obtained, which is used as the real label for supervised learning. Then, using a fixed window strategy, the long-time single-channel historical vibration signal and its strictly time-aligned real speed curve are synchronously segmented into fixed time dimensions. In this embodiment, a fixed-length sample pair (corresponding to 5s data points at the actual sampling rate) is set. After the segmented vibration segments are preprocessed by removing the mean and standardizing, a training set and a validation set are divided to construct a data foundation that meets the requirements for batch training of deep networks.

[0023] Step 2: Construct an end-to-end deep network model for the gearbox input shaft of a one-dimensional non-stationary time-series signal. Considering the strong non-stationarity and complex modulation characteristics of the gearbox vibration signal under frequent load and speed variations, this embodiment preferably employs a deep architecture with strong long-sequence modeling capabilities, such as a bidirectional state-space model or a long short-term memory network. Such architectures can efficiently capture global temporal dependencies in long vibration sequences with extremely low time complexity, robustly extracting speed characteristics from the original waveform rich in noise and impact components. During the training phase, the fixed-length sample pairs constructed in Step 1 are input into the network in batches. The mean square error between the 5-second predicted speed sequence output by the network and the true label sequence is used to construct the loss function. An optimization algorithm is used for backpropagation and parameter updates. After the model reaches convergence on the validation set, the network weights are saved and solidified, serving as the base model for subsequent deployment in actual industrial applications.

[0024] Step 3: At the edge of the actual industrial deployment, receive the raw long-term online monitoring signals continuously collected by the hoisting mechanism of the quay crane. Use an edge detection algorithm to extract the falling edge timestamps from the pulse sequence, denoted as a timestamp set. Indicates the first The trigger time of each pulse. To eliminate invalid non-working segments, the time interval between adjacent pulses is calculated. .like If the device is found to be in a stopped state, the continuous monitoring record is divided into several independent segments. All segments are iterated over, and only valid operation segments with a duration greater than a set threshold are retained. These segments are then used as the valid variable-length input sequence to be processed by the edge-side base model. (Assume its total length is) Meanwhile, a linear interpolation method is used to map the sparse rotational speed sequence onto the time axis of the vibration signal, obtaining a continuous actual rotational speed sequence aligned with the original signal time as a verification benchmark.

[0025] Step 4: Preload the converged end-to-end network model of the gearbox input shaft trained in Step 2 into the edge computing device. This is for the arbitrary variable-length input sequence extracted in Step 3. Set the parameters for the overlapping sliding window during the inference phase: set the sliding window length... The data point length is set to 5 seconds, consistent with the model input dimension. To ensure a smooth reconstructed rotational speed curve, eliminate stitching artifacts, and simultaneously consider computational efficiency at the edges, an overlap rate is set. It is 20%, which is the sliding step size. The data point length is set to 4 seconds.

[0026] Step 5: The program automatically processes the variable-length vibration sequence obtained in Step 3 according to the rules of 4-second step size and 5-second window length set in Step 4. Perform real-time slicing. Overlapping sliding cuts are performed on the complete input tensor along the time dimension. A local observation window Defined as: This operation transforms a continuous, non-stationary, variable-length data stream into a fixed-length observation subsequence containing multiple locally overlapping segments, and encapsulates it into a fixed-length sequence tensor batch suitable for batch processing by deep learning models.

[0027] Step 6: Determine if the sliding window has reached the end of the monitoring sequence. If the remaining data length at the end of the sequence is less than the set window length... (i.e., less than 5 seconds), to avoid artifacts introduced by conventional padding, an end-back processing strategy is adopted. This forces the sequence to use the last end... A complete 5-second segment is extracted from the starting point and used as the last input sample in this batch. : in, This is the maximum index of the local window obtained from the decomposition.

[0028] Step 7: The batches of sequence tensors obtained after decomposition in Steps 5 and 6 are fed into the end-to-end speed estimation model in the edge computing device for parallel computation to obtain the local instantaneous speed prediction results for each 5-second segment, thus obtaining the corresponding local speed estimation sequence set. .

[0029] Step 8: In the reconstruction module, an overlapping summation mechanism is used to integrate and smooth the local rotational speed estimation results. Alignment is performed based on the actual positions of each segment on the original time axis. For time points in overlapping regions (e.g., based on the parameters mentioned above, the last 1 second of the first window and the first 1 second of the second window overlap), a weighted average formula is used for calculation. The denominator automatically records the number of times this time point is covered (usually twice), thus smoothing the transition boundary data and eliminating the Gibbs effect (artifacts) at the splicing points. After the window data extracted from the end-of-line backtracking is smoothly fused with the preceding window, the final output is a complete and smooth rotational speed curve of the entire start-stop process of the quay crane hoisting mechanism, strictly aligned with the time axis of the original variable-length vibration signal. Combined with this global continuous rotational speed curve, the original non-stationary vibration signal is resampled to the angular domain for subsequent order tracking and state assessment analysis.

[0030] Figures 3 to 5 This paper presents the comprehensive results of the verification process of the method of the present invention on selected test samples. It includes the original time-domain vibration signal, the actual rotational speed, the rotational speed estimated by the method, and the order spectrum analysis results. It can be seen that the original vibration signals of the test samples all exhibit strong non-stationary characteristics, characterized by amplitude modulation induced by changing operating conditions and irregular impact components. Correspondingly, the rotational speed comparison curves show that the equipment experienced drastic transient changes, including short-term acceleration / deceleration and frequent adjustment processes. However, the global instantaneous rotational speed reconstructed by the adaptive inference of the present invention still maintains a high degree of consistency with the measured reference rotational speed, effectively avoiding phase abrupt changes and hysteresis caused by hard splicing. Furthermore, after resampling the non-stationary vibration signal to the angular domain using the reconstructed rotational speed, the order spectrum exhibits sharp energy focusing characteristics at the meshing order. The spectral peaks derived from the predicted values ​​of the method are precisely aligned with the hardware measurement benchmark, further confirming the spectral fidelity and engineering practical value of the reconstructed signal.

Claims

1. A method for adaptive inference of end-to-end speed of a gearbox input shaft based on a variable-length sequence, characterized in that, Includes the following steps: Step 1: During the offline model training phase, collect long-term single-channel vibration signals of the gearbox input shaft and their aligned true rotational speed curves; Step 2: Using a fixed window strategy, the long-time single-channel vibration signal and its aligned true rotational speed curve obtained in Step 1 are synchronously divided into time-divided segments. Fixed-length sample pairs; preprocessing such as mean removal and Z-score standardization is performed on the segmented vibration segments to finally divide the training set and validation set, and construct data that meets the batch training requirements of deep learning networks; Step 3: Based on the data from Step 2, construct an end-to-end deep network model for the input shaft of the gearbox oriented towards one-dimensional non-stationary time-series signals; Step 4: During the online inference phase, obtain the variable-length continuous vibration tensor to be monitored; define a sliding window mechanism with overlap ratio to transfer the original variable-length continuous vibration tensor... This is decomposed into a series of fixed-length observation subsequences suitable for the end-to-end depth network model input of the gearbox input shaft in step 3; assuming the time length of the sliding window remains consistent with the model input constraints, i.e., The sliding step size is And set the overlap rate between adjacent windows. Then the step size satisfies ; Perform overlapping sliding truncation on the complete input tensor along the time dimension to obtain the first... Local fixed-length observation subsequence Defined as: Step 5: Determine if the sliding window has reached the physical end of the monitored data sequence; if the sliding window reaches the end and the remaining time step length is insufficient for the fixed input length required by the model. To avoid introducing non-physical high-frequency noise during conventional zero-padding, a backtracking truncation strategy is adopted: forcing the original sequence to use its physical end time point. Starting from the endpoint, extract a complete segment of length in reverse order. The data segment is used as the last subsequence, i.e., the tail subsequence, in this batch decomposition. Formal definition: in, This represents the maximum index value of the local subsequence generated after sequence decomposition. Step 6: Combine all the subsequences generated in step 4 with the tail subsequence generated in step 5 through compensation. Merge the sequences to form a complete set of subsequences to be inferred; then feed this set into the end-to-end speed estimation model of the gearbox input shaft described in step 3 for forward propagation to obtain a set of local speed estimation sequences corresponding to each local observation window. ; Step 7: To eliminate splicing artifacts at window boundaries when the model independently processes each local sequence, a weighted averaging mechanism based on coverage counts is used to integrate and smooth the local prediction results; the formal description of the global adaptive reconstruction process is as follows: in, For global refactoring at a specific point in time The final transient speed prediction value; For the first Each local window corresponds to the output at a specific time point. The predicted local rotational speed; As an indicator function, when time point In the first When within the coverage area of ​​a local window, Otherwise, it is 0; Step 8: Output the reconstructed global smooth speed curve The continuous speed curve is used to perform order tracking analysis and condition monitoring of non-stationary vibration signals of industrial machinery, achieving seamless speed estimation of continuous monitoring data of arbitrary length.

2. The method for adaptive inference of end-to-end speed of gearbox input shaft based on variable-length sequence according to claim 1, characterized in that, Step 1 involves the following steps: using an accelerometer and a photoelectric speed sensor installed near the input shaft bearing housing or housing of the gearbox, historical operating data of the electromechanical equipment under various non-stationary operating conditions are collected synchronously; the collected discrete acceleration data is converted into a one-dimensional long-time single-channel vibration signal; simultaneously, by calculating the high-frequency pulse signal recorded by the photoelectric speed sensor, a high-precision transient actual speed curve of the equipment is obtained, which is used as a real label for supervised learning.

3. The method for adaptive inference of end-to-end speed of gearbox input shaft based on variable-length sequence according to claim 1, characterized in that, In step 3, a deep architecture with strong long sequence modeling capabilities is adopted, such as a bidirectional state-space model or a long short-term memory network.

Images