Rear main reducer assembly offline self-learning detection method and system
By combining multimodal sensors and CNN-HMM models, the problems of low efficiency and insufficient single-modal data in traditional detection methods are solved, enabling efficient and accurate condition assessment and fault early warning of the rear main reducer assembly.
Patent Information
- Application Number
- CN202510945240.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-09
- Publication Date
- 2025-11-11
AI Technical Summary
Traditional methods for testing rear main reducer assemblies are inefficient, rely on manual experience, and data from single-mode sensors cannot fully reflect the performance status, leading to the omission of potential defects and a lack of dynamic adaptive capabilities.
Data is collected using multimodal sensors, and feature fusion and state assessment are performed using a CNN-HMM model. Vibration, noise, torque, temperature and image data features are extracted by CNN, and weighted fusion is performed using an attention mechanism, combined with the HMM model for state assessment.
It enables 360-degree all-round monitoring of the rear main reducer assembly, improves the accuracy of complex fault identification, enhances the precision and robustness of condition assessment, and realizes early warning of faults.
Smart Images
Figure CN120927286A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of automotive reducer production and testing technology, and particularly relates to a self-learning testing method and system for the off-line testing of a rear main reducer assembly. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] With the development of the automotive industry, the requirements for vehicle performance and safety are constantly increasing. As a core component of the vehicle's transmission system, the rear final drive (RTD) assembly requires rigorous off-line testing to ensure product quality, particularly reliability and performance stability. Traditional testing methods typically rely on manual experience and standardized automated testing equipment or single-mode sensor data. Therefore, these methods often have the following limitations: manual testing is inefficient, requiring experienced personnel to perform individual tests, and is highly subjective, making consistency difficult to guarantee; secondly, while single-mode sensors are commonly used in RTD off-line testing, their data cannot comprehensively reflect the performance status of the RTD assembly, potentially leading to the omission of potential defects.
[0004] In recent years, with the development of deep learning technology, data-driven intelligent detection methods have been gradually applied in industrial fields. For example, convolutional neural networks (CNNs) perform well in image recognition and time-series signal analysis, while hidden Markov models (HMMs) are adept at handling temporal state changes. However, existing research often focuses on single-modal data (such as vibration or noise analysis only), failing to fully utilize the complementarity of multi-sensor data, resulting in limited detection accuracy. Furthermore, traditional methods lack dynamic adaptive capabilities and cannot achieve real-time state assessment and fault self-learning in production line environments. Summary of the Invention
[0005] To overcome the shortcomings of the prior art, this invention provides a self-learning detection method and system for the off-line main reducer assembly. Its core lies in collecting multi-dimensional data through multi-modal sensors and using a CNN-HMM model for feature fusion and state assessment, ultimately achieving intelligent detection and fault location.
[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a self-learning detection method for the off-line assembly of a rear main reducer; A self-learning detection method for a rear main reducer assembly after production line completion includes: The rear main reducer assembly to be tested is fixed on the test bench, and the actual working conditions are simulated by the speed control unit. The operating data of the rear main reducer assembly is collected by a multi-modal sensor module and preprocessed. The preprocessed running data is transmitted to a CNN-HMM model to evaluate the operating status of the rear main reducer assembly in real time. Specifically, CNN is used to extract features from the preprocessed vibration signal data, noise signal data, torque, temperature data, and image data. An attention mechanism is used to weight and fuse the extracted features to obtain a comprehensive feature vector. This comprehensive feature vector is then input into the HMM module to obtain the operating status sequence of the rear main reducer assembly and the confidence rate of the current operating status. The operating status of the rear main reducer assembly is then evaluated based on the confidence rate. The real-time evaluation results are compared with preset standards to generate an inspection report. Based on the inspection results, fault diagnosis is performed on non-conforming products and maintenance suggestions are generated.
[0007] As a further technical solution, the multimodal sensor module includes a vibration sensor, a noise sensor, a torque sensor, a temperature sensor, and a vision sensor; As a further technical solution, the operating data includes vibration signal data, noise signal data, torque, temperature data, and image data.
[0008] As a further technical solution, the preprocessing includes filtering and normalization.
[0009] As a further technical solution, CNN is used to extract features from vibration signal data, noise signal data, torque, temperature data, and image data, including: The vibration signal data is subjected to short-time Fourier transform to generate a time-frequency graph, and the time-frequency graph is convolved to obtain the vibration signal features; Noise in the noise signal data is removed by bandpass filtering, and a Mel spectrogram is generated; the Mel spectrogram is then input into the convolutional layer of a CNN to extract noise signal features. The torque signal data is synchronized with vibration and noise signals. The temporal torque signal is converted into a two-dimensional image using Gram angle field, and torque features are extracted by convolution. Heat maps are generated based on temperature data, and global temperature distribution is extracted using CNN to obtain temperature features; The image data is converted into RGB images, and image features are extracted using convolutional layers.
[0010] As a further technical solution, an attention mechanism is used to perform weighted fusion of the extracted features to obtain a comprehensive feature vector, including: The extracted vibration signal features, noise signal features, torque features, temperature features, and image features are mapped to a unified dimension through a fully connected layer; Calculate the weight vector using a query-key mechanism; The weight vector is assigned to each modality feature, and the comprehensive feature vector is obtained by weighted summation.
[0011] As a further technical solution, the comprehensive feature vector is input into the HMM module to obtain the operating state sequence of the rear main reducer assembly and the confidence rate of the current operating state. Based on the confidence rate, the operating state of the rear main reducer assembly is evaluated, including: The Viterbi algorithm is used for state definition and decoding to calculate the most likely hidden state sequence corresponding to the current feature vector. The probability distribution of each state at the current time is calculated using a forward-backward algorithm, and the maximum probability is taken as the confidence rate of the current state. Set a confidence rate threshold and evaluate the operating status of the rear main reducer assembly based on the threshold.
[0012] The second aspect of the present invention provides a self-learning detection system for the off-line production of a rear main reducer assembly.
[0013] A self-learning detection system for the off-line production of a rear main reducer assembly, comprising: The multimodal data acquisition module is configured to: fix the rear main reducer assembly to be tested on the test bench and simulate the actual working conditions through the speed control unit; The data preprocessing module is configured to: collect the operating data of the rear main reducer assembly through the multimodal sensor module and perform preprocessing; The status assessment module is configured to transmit the preprocessed running data to the CNN-HMM model to perform real-time assessment of the running status of the rear main reducer assembly; wherein, CNN is used to extract features from the preprocessed vibration signal data, noise signal data, torque, temperature data and image data respectively. The extracted features are weighted and fused using an attention mechanism to obtain a comprehensive feature vector. The integrated feature vector is input into the HMM module to obtain the operating state sequence of the rear main reducer assembly and the confidence rate of the current operating state. The operating state of the rear main reducer assembly is evaluated based on the confidence rate. The fault detection and analysis module is configured to: compare real-time evaluation results with preset standards, generate a test report, and diagnose faults in non-conforming products and generate maintenance suggestions based on the test results.
[0014] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a self-learning detection method for the offline production of a rear main reducer assembly as described in the first aspect of the present invention.
[0015] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a self-learning detection method for a rear main reducer assembly as described in the first aspect of the present invention.
[0016] The above one or more technical solutions have the following beneficial effects: (1) By integrating multimodal data of vibration, noise, torque, temperature and vision, this invention breaks through the limitations of traditional single sensor detection and realizes 360-degree all-round monitoring of the operating status of the rear main reducer assembly. The data of each sensor corroborate each other, which can effectively avoid misjudgment and missed detection, and greatly improve the accuracy of identifying complex faults (such as the simultaneous presence of gear wear and bearing loosening).
[0017] (2) This invention uses a CNN model to extract deep features from data of each modality and combines it with an attention mechanism for dynamic weighted fusion, which can adaptively highlight key fault features and effectively improve the accuracy and robustness of condition assessment. Based on the HMM model, the analysis of time-series features can accurately identify the operating status change trend of the rear main reducer assembly, and the assessment results can be quantified by confidence rate to achieve early warning of faults.
[0018] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0020] Figure 1 This is a flowchart of the method in the first embodiment.
[0021] Figure 2 This is a system structure diagram of the second embodiment. Detailed Implementation
[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0023] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.
[0024] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0025] Example 1 This embodiment discloses a self-learning detection method for the off-line main reducer assembly; First, adjust the fixing device of the testing bench according to the model and size of the rear main reducer assembly to ensure it is securely installed. Start the speed control unit, set a suitable speed range, and simulate the actual operating conditions of the rear main reducer. Install the vibration sensor, noise sensor, torque sensor, temperature sensor, and vision sensor in their respective positions on the rear main reducer assembly and calibrate them to ensure accurate data acquisition.
[0026] Furthermore, in combination Figure 1 A self-learning detection method for the off-line assembly of a rear main reducer includes: Step S101: Fix the rear main reducer assembly to be tested on the test bench and simulate the actual working conditions through the speed control unit; Step S102: Collect the operating data of the rear main reducer assembly through the multi-modal sensor module and perform preprocessing. Step S103: The preprocessed running data is transmitted to the CNN-HMM model to evaluate the running status of the rear main reducer assembly in real time. Specifically, CNN is used to extract features from the preprocessed vibration signal data, noise signal data, torque, temperature data, and image data. An attention mechanism is used to weight and fuse the extracted features to obtain a comprehensive feature vector. This comprehensive feature vector is then input into the HMM module to obtain the running status sequence of the rear main reducer assembly and the confidence rate of the current running status. The running status of the rear main reducer assembly is then evaluated based on the confidence rate. Step S104: Compare the real-time evaluation results with preset standards to generate a test report. Based on the test results, diagnose the faults of the unqualified products and generate maintenance suggestions.
[0027] Furthermore, the above steps also include the following: In step S101, the fixing device of the testing bench is first adjusted according to the model and size of the rear main reducer assembly to ensure its secure installation. The speed control unit is started, and a suitable speed range is set to simulate the actual operating conditions of the rear main reducer. The multi-modal sensors are installed in the corresponding positions of the rear main reducer assembly and calibrated to ensure accurate acquisition of sensor data.
[0028] In step S102, the multimodal sensor module includes a vibration sensor, a noise sensor, a torque sensor, a temperature sensor, and a vision sensor.
[0029] Specifically, vibration sensors are installed in key parts of the rear main reducer assembly (such as the housing and gear shafts) to collect vibration signal data in real time and detect mechanical faults (such as gear wear and bearing damage). Noise sensors are arranged around the rear main reducer assembly to collect noise signal data during operation and analyze for abnormal noise (such as whistling and unusual noises). Torque sensors are installed on the input and output shafts to monitor torque transfer in real time and determine if there are transmission faults. Temperature sensors monitor the operating temperature of the rear main reducer assembly to prevent overheating. A vision sensor using an industrial camera captures images of the rear main reducer assembly's appearance for subsequent inspection for defects such as cracks and deformation.
[0030] Furthermore, the vibration signal data, noise signal data, torque, temperature data, and image data are preprocessed. First, the asynchronous data streams collected by each sensor are timestamped, and time alignment of multi-source data is achieved based on a sliding window mechanism. The rotational speed signal is resampled to ensure consistency with the sampling rate of other sensors. Second, for the vibration signal data, a Butterworth bandpass filter is used to remove high-frequency noise and power frequency interference, and signal normalization is performed. A time-frequency graph is generated through short-time Fourier transform. For the noise signal data, an FIR bandpass filter is used to eliminate environmental noise, and a Mel filter bank is used to generate a Mel spectrum. For the torque signal, a Gram angle field is used to convert the time-series torque signal into a two-dimensional grayscale image. Temperature data is corrected, and bilinear interpolation is used to expand discrete temperature measurement points into a 128×128 thermal map. The acquired image data is first filtered for noise using median filtering and then converted into RGB images.
[0031] In step S103, the CNN model employs five parallel branches to extract features from the preprocessed vibration signal data, noise signal data, torque, temperature data, and image data, respectively. Specifically, during vibration signal feature extraction, the time-frequency graph is input into the vibration signal feature extraction branch of the CNN model. Shallow convolutions (e.g., 3x3) are used to capture local frequency domain patterns, followed by deep convolutions (e.g., 5x5) to extract broadband features. Max pooling layers are then used to reduce dimensionality while preserving salient features. Finally, fully connected layers output the vibration signal features.
[0032] In the torque feature extraction branch, a dual-branch structure is used to process the sine and cosine components of the two-dimensional grayscale image converted from the Gram angle field. Each branch contains two residual blocks, and the torque features are obtained by fusing the features of the two branches. For temperature features, a convolutional layer is used to extract the global temperature distribution from the heatmap to obtain the temperature features.
[0033] In the image feature extraction process, the image data is converted into an RGB image. The convolutional layer uses a 3×3 filter for local feature extraction, and each layer uses the ReLU activation function to enhance nonlinearity. Next, it is downsampled through a 2×2 max pooling layer to retain salient features and reduce dimensionality. Then, high-order semantic features are extracted layer by layer by stacking multiple convolutional-pooling modules (usually 4-6 layers). The shallow layer captures low-level features such as edges and textures, while the deep layer identifies abstract features such as component structures. Finally, a fully connected layer is used to convert the feature map into a fixed-dimensional feature vector to obtain the image features.
[0034] Furthermore, the extracted vibration signal features v, noise signal features a, torque features t, temperature features h, and image features i are mapped to a unified dimension through a fully connected layer, resulting in:
[0035] Secondly, a learnable global context vector is defined. The query vector is used to evaluate the importance of each mode. Vibration signal feature v, noise signal feature a, torque feature t, temperature feature h, and image feature i are used as the key vector and value vector, respectively. The similarity between the query and the key is calculated using the dot product to obtain the score.
[0036] In the formula, For the score, This is for the transpose operation.
[0037] Then, the weights are obtained using Softmax normalization, i.e.:
[0038] In the formula, This is the weight vector.
[0039] The weight vectors are assigned to each modality feature, and the weighted sum is obtained to obtain the comprehensive feature vector, as shown in the following formula:
[0040] In the formula, This is a comprehensive feature vector.
[0041] Furthermore, the synthesized feature vector is input as the observation sequence O into the HMM module, and the Viterbi algorithm is used for state definition and decoding. Specifically, a set of hidden states is defined. , This includes the normal state. ,abnormal and faults The statistical state transition frequency is used as the state transition matrix A, such as from the normal state... to abnormal The number of times or all from the normal state The number of transitions. For each hidden state, the frequency of the observed sequence value is counted and used as the observation probability B. The proportion of each state at the initial time is counted and used as the initial state probability π.
[0042] Given the trained HMM parameters (state transition matrix A, observation probability matrix B, initial state probability π) and the sequence of comprehensive feature vectors extracted by CNN, the Viterbi algorithm aims to find the most likely hidden state sequence, that is, the complete state evolution path of the rear main reducer from normal to fault.
[0043] For each possible state Calculate its initial probability:
[0044] In the formula, Initially in state The probability, In the state The first observation is generated below. The probability of.
[0045] Record the initial path .
[0046] For each subsequent time step t and each state Calculate the path with the highest probability:
[0047] In the formula, The state of the previous moment The maximum probability; From Transferred to The probability of; For state Generate current observations The probability of.
[0048] Record the optimal predecessor state:
[0049] Finally, find the most likely state at the final moment T:
[0050] in, The state of T at the final moment.
[0051] Through the Path backtracking is performed to obtain the optimal state sequence. Further, for... Normalization is performed to obtain the current state probability. By setting a confidence rate threshold, the operating status of the rear main reducer assembly is evaluated based on the threshold.
[0052] In step S104, based on the above evaluation results, a comparison is made with preset standards to generate a test report for the rear reducer assembly. Based on the test results, fault diagnosis is performed on unqualified products and repair suggestions are generated.
[0053] Example 2 This embodiment discloses a self-learning detection system for the off-line production of a rear main reducer assembly; like Figure 2 As shown, a self-learning detection system for a rear main reducer assembly after production includes: The multimodal data acquisition module 201 is configured to: fix the rear main reducer assembly to be tested on the test bench and simulate the actual working conditions through the speed control unit; The data preprocessing module 202 is configured to: collect the operating data of the rear main reducer assembly through the multimodal sensor module and perform preprocessing; The state assessment module 203 is configured to: transmit preprocessed operating data to a CNN-HMM model to perform real-time assessment of the operating state of the rear main reducer assembly; wherein, CNN is used to extract features from the preprocessed vibration signal data, noise signal data, torque, temperature data, and image data respectively; the extracted features are weighted and fused using an attention mechanism to obtain a comprehensive feature vector; the comprehensive feature vector is input into the HMM module to obtain the operating state sequence of the rear main reducer assembly and the confidence rate of the current operating state, and the operating state of the rear main reducer assembly is assessed based on the confidence rate; The fault detection and analysis module 204 is configured to: compare the real-time evaluation results with preset standards, generate a test report, and, based on the test results, diagnose the faults of unqualified products and generate maintenance suggestions.
[0054] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.
[0055] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a self-learning detection method for the off-line production of a rear main reducer assembly as described in Embodiment 1.
[0056] Example 4 The purpose of this embodiment is to provide an electronic device.
[0057] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the self-learning detection method for a rear main reducer assembly as described in Embodiment 1.
[0058] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.
[0059] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.
[0060] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A self-learning detection method for a rear main reducer assembly after production line, characterized in that, include: The rear main reducer assembly to be tested is fixed on the test bench, and the actual working conditions are simulated by the speed control unit. The operating data of the rear main reducer assembly is collected by a multi-modal sensor module and preprocessed. The preprocessed running data is transmitted to the CNN-HMM model to evaluate the running status of the rear main reducer assembly in real time. Specifically, CNN is used to extract features from the preprocessed vibration signal data, noise signal data, torque, temperature data, and image data. The extracted features are weighted and fused using an attention mechanism to obtain a comprehensive feature vector. The integrated feature vector is input into the HMM module to obtain the operating state sequence of the rear main reducer assembly and the confidence rate of the current operating state. The operating state of the rear main reducer assembly is evaluated based on the confidence rate. The real-time evaluation results are compared with preset standards to generate an inspection report. Based on the inspection results, fault diagnosis is performed on non-conforming products and maintenance suggestions are generated.
2. The self-learning detection method for the off-line assembly of the rear main reducer as described in claim 1, characterized in that, The multimodal sensor module includes a vibration sensor, a noise sensor, a torque sensor, a temperature sensor, and a vision sensor.
3. The self-learning detection method for the off-line assembly of a rear main reducer as described in claim 1, characterized in that, The operational data includes vibration signal data, noise signal data, torque, temperature data, and image data.
4. The self-learning detection method for the off-line assembly of a rear main reducer as described in claim 1, characterized in that, The preprocessing includes filtering and normalization.
5. The self-learning detection method for the off-line assembly of a rear main reducer as described in claim 1, characterized in that, CNNs were used to extract features from vibration signal data, noise signal data, torque, temperature data, and image data, including: The vibration signal data is subjected to short-time Fourier transform to generate a time-frequency graph, and the time-frequency graph is convolved to obtain the vibration signal features; Noise in the noise signal data is removed by bandpass filtering, and a Mel spectrogram is generated; the Mel spectrogram is then input into the convolutional layer of a CNN to extract noise signal features. The torque signal data is synchronized with vibration and noise signals. The temporal torque signal is converted into a two-dimensional image using Gram angle field, and torque features are extracted by convolution. Heat maps are generated based on temperature data, and global temperature distribution is extracted using CNN to obtain temperature features; The image data is converted into RGB images, and image features are extracted using convolutional layers.
6. The self-learning detection method for the off-line assembly of a rear main reducer as described in claim 1, characterized in that, The extracted features are weighted and fused using an attention mechanism to obtain a comprehensive feature vector, including: The extracted vibration signal features, noise signal features, torque features, temperature features, and image features are mapped to a unified dimension through a fully connected layer; Calculate the weight vector using a query-key mechanism; The weight vector is assigned to each modality feature, and the comprehensive feature vector is obtained by weighted summation.
7. The self-learning detection method for the off-line assembly of a rear main reducer as described in claim 1, characterized in that, The comprehensive feature vector is input into the HMM module to obtain the operating state sequence of the rear main reducer assembly and the confidence rate of the current operating state. Based on the confidence rate, the operating state of the rear main reducer assembly is evaluated, including: The Viterbi algorithm is used for state definition and decoding to calculate the most likely hidden state sequence corresponding to the current feature vector. The probability distribution of each state at the current time is calculated using a forward-backward algorithm, and the maximum probability is taken as the confidence rate of the current state. Set a confidence rate threshold and evaluate the operating status of the rear main reducer assembly based on the threshold.
8. A self-learning detection system for a rear main reducer assembly after production, characterized in that: include: The multimodal data acquisition module is configured to: fix the rear main reducer assembly to be tested on the test bench and simulate the actual working conditions through the speed control unit; The data preprocessing module is configured to: collect the operating data of the rear main reducer assembly through the multimodal sensor module and perform preprocessing; The state assessment module is configured to: transmit preprocessed operating data to a CNN-HMM model to perform real-time assessment of the operating state of the rear main reducer assembly; wherein, CNN is used to extract features from the preprocessed vibration signal data, noise signal data, torque, temperature data, and image data respectively; the extracted features are weighted and fused using an attention mechanism to obtain a comprehensive feature vector; the comprehensive feature vector is input into the HMM module to obtain the operating state sequence of the rear main reducer assembly and the confidence rate of the current operating state, and the operating state of the rear main reducer assembly is assessed based on the confidence rate; The fault detection and analysis module is configured to: compare real-time evaluation results with preset standards, generate a test report, and diagnose faults in non-conforming products and generate maintenance suggestions based on the test results.
9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the self-learning detection method for the offline production of a rear main reducer assembly as described in any one of claims 1-7.
10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the self-learning detection method for the off-line production of a rear main reducer assembly as described in any one of claims 1-7.