Small sample fault diagnosis method based on hyper-dimensional computation
By employing hyperdimensional computing technology, combined with local feature extraction and similarity measurement standards, the problem of training deep learning models in mechanical equipment fault diagnosis, which relies on large amounts of data and is costly, is solved. This enables low-cost and efficient fault diagnosis, improving the accuracy and robustness of fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YANSHAN UNIV
- Filing Date
- 2023-11-14
- Publication Date
- 2026-06-26
AI Technical Summary
Existing deep learning models for mechanical equipment fault diagnosis suffer from problems such as reliance on big data for training, large number of parameters, lack of universality, and high training costs, making it difficult to efficiently and accurately identify complex mechanical equipment fault characteristics.
By employing hyperdimensional computing technology, the vibration signals of mechanical equipment are transformed into low-dimensional feature vectors through local feature extraction and efficient preprocessing. The hyperdimensional computing model is then used for small-sample training, and a similarity measurement standard is designed for fault diagnosis.
It achieves low-cost and efficient fault diagnosis, improves the accuracy and robustness of fault diagnosis, expands the application scope of the model, reduces training costs, and enhances the model's versatility.
Smart Images
Figure CN117370762B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of fault diagnosis technology for mechanical equipment, and in particular relates to an efficient fault diagnosis method based on hyperdimensional computing. Background Technology
[0002] With the increasing level of new industrialization and the rapid development of artificial intelligence technology, industrial systems are gradually expanding in scale, and mechanical equipment is becoming increasingly precise, diversified, and automated, bringing numerous conveniences to people's daily lives and social production. The digitalization, intelligentization, and networking of the industrial economy have become the core development direction of the Fourth Industrial Revolution. To ensure the reliable operation of mechanical equipment, reduce maintenance costs, and effectively prevent and avoid safety accidents, conducting research on mechanical equipment fault diagnosis has become an important issue in the manufacturing field. Mechanical equipment typically consists of multiple components, and in actual operation, there are complex coupling relationships between these components, resulting in various state signals that are usually nonlinear and contain a large amount of noise, making fault characteristic identification of mechanical equipment difficult.
[0003] Scholars both domestically and internationally have proposed numerous methods for fault diagnosis of mechanical equipment, especially those based on machine learning and deep learning technologies. Compared to traditional signal processing-based methods, deep learning can handle more complex fault signals, mine deeper information from data, and possesses powerful feature learning and nonlinear system modeling capabilities. However, deep learning models are massive in scale, requiring training based on large datasets, parameter-free parameter tuning, and floating-point iterative operations; the number of parameters needed for training typically exceeds one million. Furthermore, different network structures are often required for different fault problems, necessitating model retraining and resulting in a lack of model universality.
[0004] Hyperdimensional computing (HDC) is an emerging cognitive model inspired by the working mechanisms of the brain. It primarily processes high-dimensional, stochastic, holographic, distributed representations of information, offering advantages such as low-cost computation, single-cycle training, and independence from large-scale data processing. This technology has already found numerous applications in speech recognition, language recognition, and handwritten digit recognition. Therefore, to more efficiently diagnose mechanical equipment faults and improve their accuracy and reliability, it is necessary to apply HDC technology to the field of fault diagnosis and design a low-cost, efficient, and universal equipment fault diagnosis technique. Summary of the Invention
[0005] The purpose of this invention is to provide a small-sample fault diagnosis method based on hyperdimensional computation, which can efficiently extract signal features of mechanical equipment and improve the performance and robustness of mechanical equipment fault diagnosis.
[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:
[0007] A small-sample fault diagnosis method based on hyperdimensional computation includes the following steps:
[0008] Step 1: Collect raw vibration signal data from various mechanical equipment, design an efficient signal preprocessing module, and extract local features from the raw vibration signals;
[0009] Step 2: Design a time series sample encoding method for hyperdimensional computation, namely multivariate sample encoding, to encode the extracted local features of the vibration signal into a series of hyperdimensional vectors, forming a hyperdimensional space;
[0010] Step 3: Determine the similarity measurement standard, design the similarity measurement fusion decision, input the local feature information of the vibration signal into the hyperdimensional calculation model for training, and the training result is called the query hypervector;
[0011] Step 4: Test and classify the trained model using test sets of different proportions to obtain the final result of model training, and analyze and verify the model using multiple fault data.
[0012] A further improvement to the technical solution of the present invention is that step 1 includes the following specific steps:
[0013] Step 5: Label the raw vibration data of the multiple mechanical devices according to the number of fault types;
[0014] Step 6: Use a sliding window to divide the original data into n equal segments of length L;
[0015] Step 7: Select the basic time-domain and frequency-domain features of the fault signal to be extracted, including root mean square, variance, maximum value, skewness, kurtosis, peak-to-peak value, spectral skewness, spectral kurtosis and power spectrum;
[0016] Step 8: Extract local feature information from the divided data segments and calculate the basic time-frequency feature information of each data segment;
[0017] Step 9: Concatenate the feature information obtained from the same original data segment to obtain the dataset used for network training.
[0018] A further improvement to the technical solution of the present invention is that step 2 includes the following specific steps:
[0019] Step 10: Design an address encoding method for hyperdimensional computation, perform nonlinear mapping on the extracted local feature information to obtain hyperdimensional address vectors. The vectors are independent and unique, and the purpose is to project the feature vectors onto the high-dimensional hyperdimensional vectors.
[0020] Step 11: Design a numerical encoding method for hyperdimensional computation. Encode the amplitude of local signal features using a continuous mapping method to obtain a numerical hyperdimensional vector:
[0021] (1) Randomly select a hyperdimensional vector from the d-dimensional space to represent the minimum value MIN in the interval;
[0022] (2) Randomly flip d / 2 / (L-1) bits in the hyperdimensional vector corresponding to the previous value to encode the next value. Each bit can only be flipped once and is not flipped repeatedly.
[0023] (3) Repeat step 2 until all L values are encoded, and finally obtain L numerical hyperdimensional vectors.
[0024] Step 12: Design a sample encoding method for hyperdimensional computation. Use a multivariate sample encoding method to encode the time series data. The encoded hyperdimensional vector will uniquely identify the sample, represent all the features of the sample, and serve as the input for hyperdimensional computation.
[0025] A further improvement to the technical solution of the present invention is that step 3 includes the following specific steps:
[0026] Step 13: Store the encoded address and numerical superdimensional vector in a storage space called the project memory, so that it can be used as raw material for subsequent algorithms;
[0027] Step 14: Standardize the training samples obtained after encoding the samples to obtain the training hyperdimensional vector;
[0028] Step 15: Combine the training hypervectors belonging to the same class into a set through addition, and then standardize them to obtain the class hypervectors;
[0029] Step 16: The category hypervectors are stored together with the category labels in a memory space called associative memory, completing the training.
[0030] A further improvement to the technical solution of the present invention is that step 4 includes the following specific steps:
[0031] Step 17: Encode the test sample using the address hyperdimensional vector and numerical hyperdimensional vector in the project memory, using the same encoding method as the training sample. The encoding result is called the query hyperdimensional vector.
[0032] Step 18: Determine the similarity measurement method and design a similarity measurement fusion decision. Calculate the similarity between the query hyperdimensional vector and the category hyperdimensional vector in the associative memory. Similarity measurement methods include Hamming distance, cosine distance, dot product, and Jaccard distance. The calculation methods are as follows:
[0033] Hamming distance:
[0034]
[0035] Where d is the dimension of the hyperspace. When the normalized Hamming distance between two hyperdimensional vectors is 0.5, the two hyperdimensional vectors are (pseudo) orthogonal to each other. The closer the normalized Hamming distance is to 0, the better the similarity between the two hyperdimensional vectors. Conversely, the closer the normalized Hamming distance is to 1, the better the orthogonality and the lower the similarity.
[0036] Cosine distance:
[0037]
[0038] The closer the cosine distance between hyperdimensional vectors is to 0, the higher their orthogonality; the closer it is to 1, the better their similarity; when all corresponding elements are different, the cosine distance is -1.
[0039] Dot product:
[0040]
[0041] A larger dot product value indicates a better similarity between the elements, and vice versa.
[0042] Jaccard distance:
[0043]
[0044] Here, ∧ represents logical AND and ∨ represents logical OR. The larger the value, the better the similarity between them, and vice versa.
[0045] Step 19: Determine the category corresponding to the highest similarity obtained from different similarity metrics. The category with the highest number of similarities is the category to which the sample belongs in the hyperdimensional calculation. For example, if Hamming distance classifies the sample into category 1, cosine distance into category 2, dot product into category 3, and Jaccard distance into category 1, then the sample will ultimately be classified into category 1. If the sample is classified into different categories by the four distances, then the category classified by cosine distance will be selected as the final category.
[0046] Step 20: Analyze and validate the obtained model using multiple fault datasets, including bearing fault datasets, gearbox fault datasets, and composite fault datasets, to obtain the final fault diagnosis system.
[0047] The technological advancement achieved by this invention is that, by using manually selected features to extract local features from the original vibration signal data, the length of the extracted feature information is much shorter than that of the original fault data, thus reducing the training load of the model.
[0048] By proposing an intelligent industrial equipment fault diagnosis method and system based on hyperdimensional computing, hyperdimensional computing is applied to the field of mechanical equipment fault diagnosis for the first time. Moreover, by combining the time-frequency characteristics of the fault signal itself, the complex high-dimensional raw data is transformed into a lower-dimensional feature vector through an efficient preprocessing method, which reduces the amount of data processing in the model and makes the model more efficient.
[0049] By combining the advantages of hyperdimensional computing models, the model can achieve excellent performance with only a small number of samples in a single training run, which greatly speeds up the model training process and reduces the cost of model training.
[0050] Multiple fault datasets were created to validate the generality of the hyperdimensional computation model, expanding its application scope.
[0051] Therefore, this invention can extract more efficient fault diagnosis information, enhance the performance and accuracy of fault diagnosis, and has strong versatility, providing a new technology for fault diagnosis systems of mechanical equipment. Attached Figure Description
[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0053] Figure 1 This is a flowchart of the method of the present invention;
[0054] Figure 2 This is a flowchart for extracting local features from vibration signals;
[0055] Figure 3 This is a flowchart of the sample encoding process;
[0056] Figure 4 This is the overall flowchart for training and testing hyperdimensional computing models. Detailed Implementation
[0057] The technical solution of the present invention will be further described in detail below with reference to the embodiments:
[0058] Example 1
[0059] A bearing failure simulation test bench was used, with motor loads ranging from 0 hp to 3 hp, a signal sampling frequency of 12 kHz, and bearing failure diameters ranging from 0.007 inches to 0.028 inches. Failure types were categorized into three types: inner race failure, outer race failure, and rolling element failure. Each operating condition included ten different state types and sizes of data, including a health dataset.
[0060] Intelligent industrial equipment fault diagnosis method and system process based on hyperdimensional computing, as follows: Figure 1 As shown,
[0061] Step 1: Collect raw vibration signal data from various mechanical devices, design an efficient signal preprocessing module, and extract local features from the raw vibration signals, including the following steps:
[0062] Step 1.1: Label the raw vibration data of the multiple mechanical devices according to the number of fault types;
[0063] Step 1.2: Use a sliding window to divide the original data (length 512) into 10 equal segments of length 51;
[0064] Step 1.3: Select the basic time-domain and frequency-domain features of the fault signal to be extracted, including root mean square, variance, maximum value, skewness, kurtosis, peak-to-peak value, spectral skewness, spectral kurtosis and power spectrum;
[0065] Step 1.4: Extract local feature information from the divided data segments and calculate the basic time-frequency feature information of each data segment;
[0066] Step 1.5: Concatenate the feature information obtained from the same original data segment to obtain the dataset used for training the hyperdimensional network.
[0067] Step 2: Design a time-series sample encoding method for hyperdimensional computation, namely multivariate sample encoding, to encode the extracted local features of the vibration signal into a series of hyperdimensional vectors, forming a hyperdimensional space. This includes the following steps:
[0068] Step 2.1: Design an address encoding method for hyperdimensional computation, perform nonlinear mapping on the extracted local feature information to obtain hyperdimensional address vectors. The vectors are independent and unique, and the purpose is to project the feature vectors onto the high-dimensional hyperdimensional vectors.
[0069] Step 2.2: Design a numerical encoding method for hyperdimensional computation, and encode the amplitude of local signal features based on a continuous mapping method:
[0070] (1) Randomly select a hyperdimensional vector from the d-dimensional space to represent the minimum value 0 in the interval;
[0071] (2) Randomly flip the d / 2 / 100 bits of the hyperdimensional vector corresponding to the previous value to encode the next value. Each bit can only be flipped once and is not flipped repeatedly.
[0072] (3) Repeat step 2 until all 100 values are encoded, and finally obtain 100 numerical hyperdimensional vectors.
[0073] Step 2.3: Design a sample encoding method for hyperdimensional computation. A multivariate sample encoding method is adopted. The encoded hyperdimensional vector will uniquely identify the sample, represent all the features of the sample, and serve as the input for hyperdimensional computation.
[0074] Step 3: Determine the similarity measurement standard, input the local feature information of the vibration signal into the hyperdimensional computational model for training, and the training result is called the query hypervector. This includes the following steps:
[0075] Step 3.1: Store the encoded address and numerical hyperdimensional vector in a storage space called the project memory, so that it can be used as raw material for subsequent algorithms;
[0076] Step 3.2: Standardize the training samples obtained after encoding the samples to obtain the training hyperdimensional vector;
[0077] Step 3.3: Combine the training hypervectors belonging to the same class into a set through addition, and then standardize them to obtain the class hypervectors;
[0078] Step 3.4: The category hypervectors are stored together with the category labels in a memory space called associative memory, completing the training.
[0079] Step 4: Determine the similarity measurement standard, design a similarity measurement fusion decision, test the trained model using test sets of different proportions to obtain the final result of model training, and analyze and verify the model using multiple fault data sets. This includes the following steps:
[0080] Step 4.1: Encode the test sample using the address hyperdimensional vector and numerical hyperdimensional vector in the project memory, using the same encoding method as the training sample. The encoding result is called the query hyperdimensional vector.
[0081] Step 4.2: Design a similarity measurement method and a decision fusion method to calculate the similarity between the query hyperdimensional vector and the category hyperdimensional vector in the associative memory, respectively. The similarity measurement method is as follows:
[0082] Hamming distance:
[0083]
[0084] Where d is the dimension of the hyperspace. When the normalized Hamming distance between two hyperdimensional vectors is 0.5, the two hyperdimensional vectors are (pseudo) orthogonal to each other. The closer the normalized Hamming distance is to 0, the better the similarity between the two hyperdimensional vectors. Conversely, the closer the normalized Hamming distance is to 1, the better the orthogonality and the lower the similarity.
[0085] Cosine distance:
[0086]
[0087] The closer the cosine distance between hyperdimensional vectors is to 0, the higher their orthogonality; the closer it is to 1, the better their similarity; when all corresponding elements are different, the cosine distance is -1.
[0088] Dot product:
[0089]
[0090] A larger dot product value indicates a better similarity between the elements, and vice versa.
[0091] Jaccard distance:
[0092]
[0093] Here, ∧ represents logical AND and ∨ represents logical OR. The larger the value, the better the similarity between them, and vice versa.
[0094] Step 4.3: Determine the category corresponding to the highest similarity obtained from different similarity metrics. The category with the highest number of similarities is the category to which the sample belongs in the hyperdimensional calculation. For example, if Hamming distance classifies the sample into category 1, cosine distance into category 2, dot product into category 3, and Jaccard distance into category 1, then the sample will ultimately be classified into category 1. If the sample is classified into different categories by the four distances, then the category classified by cosine distance will be selected as the final category.
[0095] Step 4.4: Analyze and verify the obtained model using the fault test dataset to obtain the final fault diagnosis system.
[0096] This embodiment uses the following comparison method:
[0097] K-Nearest Neighbors Algorithm: After local feature extraction, the K-nearest neighbor algorithm is used for model training and classification of the original vibration signal;
[0098] Neural Network: After local feature extraction, the original vibration signal is trained and classified using a neural network with two hidden layers;
[0099] Autoencoder network: After local feature extraction, the original vibration signal is trained and classified using a simple autoencoder network.
[0100] The accuracy rates of the proposed method and the control method are shown in Table 1. All results are the average of 10 random training runs. The proposed method performs well under all four operating conditions, demonstrating that the intelligent industrial equipment fault diagnosis method and system based on hyperdimensional computing can efficiently utilize the relevant features of faults extracted from vibration signals and can effectively diagnose and predict bearing faults.
[0101] Table 1. Comparison of test results of the method of the present invention with those of the control method (training set ratio 30%)
[0102]
[0103]
[0104] Training set ratio 20%
[0105]
[0106] Training set ratio 10%
[0107]
[0108] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A small-sample fault diagnosis method based on hyperdimensional computation, characterized in that... Includes the following steps: Step 1: Collect raw vibration signal data from various mechanical devices, design a signal preprocessing module, and calculate local feature information of the raw vibration signals; Step 2: Design a time-series sample encoding method for hyperdimensional computation, namely multivariate sample encoding, to encode the extracted local features of the vibration signal into a series of hyperdimensional vectors, forming a hyperdimensional space; specifically as follows: Step 2.1: Design an address encoding method for hyperdimensional computation, perform nonlinear mapping on the extracted local feature information to obtain address hyperdimensional vectors. The vectors are independent and unique. Project the feature vectors onto the high-dimensional hyperdimensional vectors. Step 2.2: Design a numerical encoding method for hyperdimensional computation. Encode the amplitude of local signal features using a continuous mapping method to obtain a numerical hyperdimensional vector; the specific operation is as follows: (1) Randomly select a hyperdimensional vector from the d-dimensional space to represent the minimum value 0 in the interval; (2) Randomly flip the d / 2 / 100 bits in the hyperdimensional vector corresponding to the previous value to encode the next value. Each bit can only be flipped once and cannot be flipped repeatedly. (3) Repeat step 2 until all 100 values are encoded, and finally obtain 100 numerical hyperdimensional vectors; Step 2.3: Design a sample encoding method for hyperdimensional computation. A multivariate sample encoding method is adopted. The encoded hyperdimensional vector will uniquely identify the sample, represent all the features of the sample, and serve as the input for hyperdimensional computation. Step 3: Input the local feature information of the vibration signal into the hyperdimensional computation model for training. The training result is called the category hypervector. Step 4: Determine the similarity measurement standard, design the similarity measurement fusion decision, use test sets of different proportions to test and classify the trained model, obtain the final result of model training, and use multiple fault data to analyze and verify the model.
2. The small-sample fault diagnosis method based on hyperdimensional computation according to claim 1, characterized in that: Step 1 is as follows: Step 1.1: Label the raw vibration data of the multiple mechanical devices according to the number of fault types; Step 1.2: Use a sliding window to divide the original data of length L into n equal segments of length L / n; Step 1.3: Select the basic time-domain and frequency-domain features of the fault signal to be extracted, including root mean square, variance, maximum value, skewness, kurtosis, peak-to-peak value, spectral skewness, spectral kurtosis and power spectrum; Step 1.4: Calculate the local feature information for each segment of data to obtain the basic time-frequency feature information of each segment; Step 1.5: Concatenate the feature information obtained from the same original data segment to obtain the dataset used for training the hyperdimensional computing network.
3. The small-sample fault diagnosis method based on hyperdimensional computation according to claim 1, characterized in that, Step 3 is as follows: Step 3.1: Store the encoded address and numerical hyperdimensional vector in a storage space called the project memory, so that it can be used as raw material for subsequent algorithms; Step 3.2: Standardize the training samples obtained after encoding the samples to obtain the training hyperdimensional vector; Step 3.3: Combine the training hypervectors belonging to the same class into a set through addition, and then standardize them to obtain the class hypervectors; Step 3.4: The category hypervectors are stored together with the category labels in a memory space called associative memory, completing the training.
4. The small-sample fault diagnosis method based on hyperdimensional computation according to claim 1, characterized in that, Step 4 is as follows: Step 4.1: Encode the test sample using the address hyperdimensional vector and numerical hyperdimensional vector in the project memory, using the same encoding method as the training sample. The encoding result is called the query hyperdimensional vector. Step 4.2: Design similarity measurement methods, calculate the similarity between the query hyperdimensional vector and the category hyperdimensional vector in the associative memory respectively. Similarity measurement methods include Hamming distance, cosine distance, dot product, and Jaccard distance. Based on this, design a decision fusion classification method. Step 4.3: Determine the category corresponding to the highest similarity obtained from different similarity measurement standards. The category with the largest number of similarities is the category to which the sample belongs in the hyperdimensional calculation. Step 4.4: Analyze and verify the obtained model using multiple fault datasets, including bearing fault datasets, gearbox fault datasets, and composite fault datasets, to obtain the final fault diagnosis system.