A reciprocating compressor rapid fault diagnosis method based on sparse grid and immune mechanism
By employing a sparse mesh and immune mechanism approach, utilizing t-SNE dimensionality reduction and Manhattan distance calculation, and combining a backtracking strategy to construct a single-layer feedforward network, the problem of fault diagnosis for reciprocating compressors was solved, achieving efficient and accurate fault identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI UNIV
- Filing Date
- 2022-06-23
- Publication Date
- 2026-07-31
AI Technical Summary
Existing intelligent fault diagnosis methods for reciprocating compressors suffer from contradictions such as the lack of fault samples in offline learning and the inability of online diagnosis to learn proactively, the abundance of normal state samples and the scarcity of fault samples, the limited types of training fault samples and the large number of actual occurrence sample types, and the time-varying nature of wear faults, making it difficult to combine anomaly monitoring with fault diagnosis.
By employing sparse grids and immune mechanisms, and using t-SNE to reduce the dimensionality of the original signal, a cell generation strategy (CG) and Manhattan distance are proposed to calculate affinity. Combined with a backtracking strategy (BC) for adaptive adjustment, a single-layer feedforward neural network model is constructed for fault diagnosis.
It improves the efficiency and accuracy of fault diagnosis, reduces time complexity, enables efficient identification of small and unknown samples, and shortens the diagnosis time.
Smart Images
Figure CN115099336B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mechanical equipment fault diagnosis, specifically a rapid fault diagnosis method for reciprocating compressors based on sparse mesh and immune mechanisms. Background Technology
[0002] Reciprocating compressors, as important industrial production equipment, convert prime mover power into gas pressure. They are widely used in petroleum, chemical and other fields. Large reciprocating compressors have complex structures, many vulnerable parts, and operate continuously, resulting in a wide variety of typical failures. Some typical failure samples are scarce, such as piston rod breakage, crankshaft breakage, and connecting rod breakage. Once a failure occurs, it can lead to serious safety accidents and huge economic losses. Wear failures in reciprocating compressors are continuous time-varying failures, and quantification methods need to be developed.
[0003] Existing intelligent fault diagnosis methods suffer from the following contradictions: First, there's the contradiction between the lack of fault samples in offline learning and the inability to actively learn from fault samples obtained through online diagnosis. Current intelligent fault diagnosis methods primarily employ a one-time learning and training approach, only identifying faults during the diagnostic process and lacking online learning capabilities unless training starts from scratch again. This lack of inheritance and the inability to continuously learn from the vast prior knowledge held by experts renders methods ineffective, as they can only learn from a small number of samples. Second, there's the contradiction between an abundance of normal state samples and a scarcity of fault samples in reciprocating compressors. While there are many normal samples, fault samples are difficult to obtain, and the lack of prior knowledge hinders effective and accurate fault diagnosis. Third, there's the contradiction between the limited types of training fault samples and the diverse types of actual fault samples. The lack of accurate models to describe complex fault behaviors, coupled with the overlapping of multiple faults in actual reciprocating compressor operation, makes it difficult to obtain all fault types in a single diagnosis. Fourth, there's the contradiction between the flexible nature of wear faults in reciprocating compressors and the rigid nature of fault outcomes. Wear faults are time-varying and cumulative, making it impossible to combine anomaly monitoring with fault diagnosis for progressive online learning. These factors hinder the development of online learning-based intelligent fault diagnosis methods for large mechanical equipment such as reciprocating compressors. Therefore, there is an urgent need to study a fast fault diagnosis algorithm with strong adaptability and continuous learning capability that can be applied to reciprocating compressors. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings and deficiencies of existing technologies and provide a rapid fault diagnosis method for reciprocating compressors based on sparse grids and immune mechanisms. The proposed cellular and backtracking strategies significantly improve continuous learning performance, encode immune cells, and make adaptive adjustments, thereby enhancing the efficiency and accuracy of fault diagnosis.
[0005] To achieve the above objectives, the technical solution of the present invention includes:
[0006] A rapid fault diagnosis method for reciprocating compressors based on sparse mesh and immune mechanisms includes:
[0007] S1. The cell grid space constructed from the original signals of the reciprocating compressor is subjected to t-SNE dimensionality reduction visualization processing, which is optimally converted into two-dimensional spatial data to enhance the difference of heterogeneous samples and reduce the complexity of the immune space. A new cell generation strategy (CG) is proposed to delete duplicate cells, avoid the dimensionality explosion problem, and reduce the temporal and spatial complexity.
[0008] S2. By endowing the cell with a coding function (CE), cell affinity is calculated using Manhattan distance, simplifying existing affinity calculation methods. Secondly, a continuous learning method based on a backward cell strategy (BC) is constructed to reduce repetitive learning.
[0009] S3. An adaptive adjustment method for models based on single-layer feedforward neural networks is used to classify and cluster samples and diagnose the faults of reciprocating compressors.
[0010] The step S1, which involves constructing a superrectangular cell space and implementing a cell generation strategy (CG) for the original signal from the reciprocating compressor, specifically includes:
[0011] S1-1. Using the proposed cell strategy, constant-size hyperrectangular cells are formed in the sample space. The cell size and position are determined by the cell generation and coordinates. The generated cells are used for subsequent encoding.
[0012] S1-2, Antigen Sample: F i =[f i1 f i2 … f in Antigen coordinates CC i =[x i1 x i2 … x in ], f ik ∈[0.0001,0.9999],x ik ∈N, k=1,2,…n;
[0013] S1-3. Determine the location of the antigen in the cell:
[0014] S1-4, CC i =[2 g ·F i y = f(x) = [x] Gaussian maximum floor function;
[0015] S1-5. Describe cell coordinates using antigens: 2 g ·F i -1 <CC i ≤2 g ·F i .
[0016] The specific steps of cell coding (CE) in step S2 are as follows:
[0017] S2-1-1, Encoding the cell;
[0018] S2-1-2. Empty cells are recorded as 0 and do not participate in the calculation; only their positions are recorded.
[0019] S2-1-3. Establish sparse matrices or tensors to obtain antigen information and record known cell tags;
[0020] S2-1-4. Merge two cells with the same encoding into one cell. The new cell will overwrite the information of the old cell.
[0021] S2-1-5. Retain sufficient information about outlier samples through reasonable algebraic settings;
[0022] S2-1-6. When there is sufficient storage space or the dimensionality of the antigen is not high, tensor indexing is used to reduce time complexity.
[0023] The Manhattan distance weight calculation step in step S2 is as follows:
[0024] S2-2-1. To distinguish the contribution of each feature of an antigen to classification, the weight of each feature is calculated before calculating affinity; the purpose of increasing the weight is to maximize the affinity value of similar cells and minimize the affinity value of dissimilar cells; described as follows:
[0025]
[0026] In the formula: m and a represent different cells, and j and k represent cell types;
[0027] S2-2-2, A single-layer feedforward network training model was constructed, and the weights were updated using gradient descent with the squared loss function:
[0028]
[0029] In the formula: α represents the learning rate, and t is the number of iterations.
[0030] The steps of the continuous learning method based on backtracking strategy (BC) in step S2 are as follows:
[0031] S2-3-1, Basic Testing Process: Principles for Determining the Similarity of Unknown Antigen Types:
[0032]
[0033] In the formula: w i The weights represent the features, and ij and k represent different types of cells.
[0034] S2-2-2 To prevent infinite affinity, aff-1 is calculated in actual programming. Therefore, the basic testing process is to calculate the weighted Manhattan distance between the current cell and the new cell activated by the unknown antigen. If the calculation result is zero, the new antigen must belong to the current cell; otherwise, it will be determined that the known antigen belongs to the same category as the cell with the highest affinity.
[0035] S2-2-3. Cell generation will degenerate before the number of unknown classes matches the number of target classes. If the affinity between cells is the same, cells of known classes have a higher priority than cells of unknown classes, and cells activated first have a higher priority than cells activated later.
[0036] S2-2-4. Backtracking strategy (BC) is a bottom-up approach that can only degenerate cells from high passage number to low passage number. The time complexity of the degradation process is O(m*n), where m is the number of samples and n is the dimension.
[0037] Compared with the prior art, the present invention has the following advantages:
[0038] First, a novel cell generation strategy (CG) is proposed. High-dimensional data is adaptively and optimally degraded to a low-dimensional space using t-SNE, making it easier to construct a hyperrectangular space. Within this hyperrectangular space, cells of constant size are generated. These cells are then described using antigen samples, assigned coordinates, and used for subsequent encoding, thus reducing useless cells. This approach reduces time complexity, improves diagnostic efficiency, and avoids the dimensionality explosion problem.
[0039] By employing a cell coding strategy (CE), the computational workload for invalid empty cells is significantly reduced, solving the problem of high space complexity in existing continuous learning immune algorithms, thereby greatly improving diagnostic efficiency and shortening diagnostic time.
[0040] The training process based on Manhattan weight affinity calculation makes Manhattan distance more suitable for quickly measuring grid distance. It achieves automatic distance adjustment based on coordinate transformation to complete the cell backtracking process. Furthermore, Manhattan distance uses integer calculations, which is more efficient than Euclidean distance. In addition, Manhattan distance is equivalent to the difference between vector dimensions, and the purpose of the weights is to adjust for dimensional differences.
[0041] The continuous learning and testing process based on backtracking strategy (BC) can achieve adaptive recognition, reduce time complexity, and has efficient recognition capabilities for both small sample data and unknown sample data. Attached Figure Description
[0042] Figure 1 This is a flowchart illustrating the principle of the present invention;
[0043] Figure 2 This is a diagram of the original acceleration signal of the reciprocating compressor collected in an example of the present invention.
[0044] Figure 3 It is a pressure signal graph;
[0045] Figure 4 It is a key phase signal diagram;
[0046] Figure 5 Visualize the reciprocating compressor dataset for t-SNE processing;
[0047] Figure 6 This relates to the impact of algebra on accuracy during backtracking. Detailed Implementation
[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, for those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.
[0049] like Figure 1 As shown, a rapid fault diagnosis method for reciprocating compressors based on sparse meshes and immune mechanisms includes the following steps:
[0050] S1. The cell grid space constructed from the original signals of the reciprocating compressor is subjected to t-SNE dimensionality reduction visualization processing, which is optimally converted into two-dimensional spatial data to enhance the difference of heterogeneous samples and reduce the complexity of the immune space. A new cell generation strategy CG is proposed to delete duplicate cells, avoid the dimensionality explosion problem, and reduce the temporal and spatial complexity.
[0051] S2. By endowing cells with encoding function CE, cell affinity is calculated through Manhattan distance, simplifying existing affinity calculation methods; secondly, a continuous learning method based on backtracking strategy BC is constructed to reduce repetitive learning.
[0052] S3. An adaptive adjustment method for models based on single-layer feedforward neural networks is used to classify and cluster samples and diagnose the faults of reciprocating compressors.
[0053] The step S1, which involves constructing a super-rectangular cell space and implementing a cell generation strategy (CG) for the original signal from the reciprocating compressor, specifically includes:
[0054] S1-1. Using the proposed cell strategy, constant-size hyperrectangular cells are formed in the sample space. The cell size and position are determined by the cell generation and coordinates. The generated cells are used for subsequent encoding.
[0055] S1-2, Antigen Sample: F i =[f i1 f i2 … f in Antigen coordinates CC i =[x i1 x i2 … x in ], f ik ∈[0.0001,0.9999],x ik ∈N, k=1,2,…n;
[0056] S1-3. Determine the location of the antigen in the cell:
[0057] S1-4, CC i =[2 g ·F i y = f(x) = [x] Gaussian maximum floor function;
[0058] S1-5. Describe cell coordinates using antigens: 2 g ·F i -1 <CC i ≤2 g ·F i .
[0059] The specific steps of cell-encoding CE in step S2 are as follows:
[0060] S2-1-1, Encoding the cell;
[0061] S2-1-2. Empty cells are recorded as 0 and do not participate in the calculation; only their positions are recorded.
[0062] S2-1-3. Establish sparse matrices or tensors to obtain antigen information and record known cell tags;
[0063] S2-1-4. Merge two cells with the same encoding into one cell. The new cell will overwrite the information of the old cell.
[0064] S2-1-5. Retain sufficient information about outlier samples through reasonable algebraic settings;
[0065] S2-1-6. When there is sufficient storage space or the dimensionality of the antigen is not high, tensor indexing is used to reduce time complexity.
[0066] The Manhattan distance weight calculation step in step S2 is as follows:
[0067] S2-2-1. To distinguish the contribution of each feature of an antigen to classification, the weight of each feature is calculated before calculating affinity; the purpose of increasing the weight is to maximize the affinity value of similar cells and minimize the affinity value of dissimilar cells; described as follows:
[0068]
[0069] In the formula: m and a represent different cells, and j and k represent cell types;
[0070] S2-2-2, A single-layer feedforward network training model was constructed, and the weights were updated using gradient descent with the squared loss function:
[0071]
[0072] In the formula: α represents the learning rate, and t is the number of iterations.
[0073] The steps of the continuous learning method based on backtracking strategy BC in step S2 are as follows:
[0074] S2-3-1, Basic Testing Process: Principles for Determining the Similarity of Unknown Antigen Types:
[0075]
[0076] In the formula: w i The weights represent the features, and ij and k represent different types of cells.
[0077] S2-2-2 To prevent infinite affinity, aff-1 is calculated in actual programming. Therefore, the basic testing process is to calculate the weighted Manhattan distance between the current cell and the new cell activated by the unknown antigen. If the calculation result is zero, the new antigen must belong to the current cell; otherwise, it will be determined that the known antigen belongs to the same category as the cell with the highest affinity.
[0078] S2-2-3. Cell generation will degenerate before the number of unknown classes matches the number of target classes. If the affinity between cells is the same, cells of known classes have a higher priority than cells of unknown classes, and cells activated first have a higher priority than cells activated later.
[0079] S2-2-4. Backtracking strategy BC is a bottom-up approach that can only degenerate cells from high passage number to low passage number. The time complexity of the degradation process is O(m*n), where m is the number of samples and n is the dimension.
[0080] The algorithm training and testing processes are shown in Table 1.
[0081] Table 1 shows the immune mechanisms corresponding to the algorithms.
[0082]
[0083]
[0084] Example 1: Diagnosis of experimental data set for reciprocating compressors
[0085] The method proposed in this invention was used to diagnose valve failures in a reciprocating compressor at a natural gas company's compressor station. Three acceleration sensors, four pressure sensors, and one critical phase sensor were used to record the operating status of the reciprocating compressor. Signals were acquired at a sampling frequency of 20 kHz for four different conditions: spring loss, valve clearance, valve breakage, and normal operation. Figure 2 This is a graph of the original acceleration signal of the reciprocating compressor collected in an example of the present invention; Figure 3 It is a pressure signal graph; Figure 4 This is a key phase signal diagram.
[0086] 1. Construction of a superrectangular cellular space and cell generation strategy (CG) based on the raw signal of a reciprocating compressor:
[0087] (1) t-SNE processes the raw signal of the compressor and performs two-dimensional visualization processing, such as Figure 5 As shown, it contains 12,000 normal signals, 8,000 valve plate breakage signals, 8,000 valve plate gap signals, and 8,000 spring missing signals. Each sample contains 8 features.
[0088] (2) Generate a hyperrectangular sample space to form hyperrectangular cells of constant size. Only 8000 samples need to be randomly selected from the above data, and 4000 samples need to be randomly selected from other faulty samples for diagnosis, thereby improving training efficiency.
[0089] (3) Cell coding (CE) and assigning 0 to empty cells to reduce invalid cells.
[0090] 2. Training the diagnostic algorithm based on affinity weights calculated using Manhattan distance:
[0091] (1) Establishing a training set: Of the 20,000 samples mentioned above, 70% are selected as the training set.
[0092] (2) Manhattan affinity calculation:
[0093]
[0094] In the formula: m and a represent different cells, and j and k represent cell types.
[0095] (3) The constructed single-layer feedforward network model is used for training, and the weights are updated using the gradient descent method with the squared loss function:
[0096]
[0097] In the formula: α represents the learning rate, and t is the number of iterations.
[0098] 3. The steps of the continuous learning and testing method based on backtracking strategy (BC) are as follows:
[0099] (1) Basic testing process: Principles for determining the similarity of unknown antigen types:
[0100]
[0101] In the formula: w i The values represent feature weights, and ij and k represent different types of cells.
[0102] (2) To prevent infinite affinity, aff-1 is calculated in actual programming. Therefore, the basic testing process is to calculate the weighted Manhattan distance between the current cell and the new cell activated by the unknown antigen. If the calculation result is zero, the new antigen must belong to the current cell. Otherwise, it will be determined that the known antigen belongs to the same category as the cell with the highest affinity.
[0103] (3) Cell generation will degenerate before the number of unknown classes matches the number of target classes. If the affinity between cells is the same, cells of known classes have a higher priority than cells of unknown classes, and cells activated first have a higher priority than cells activated later.
[0104] (4) The backtracking strategy (BC) is a bottom-up approach that can only degenerate cells from high-generation cells to low-generation cells. The time complexity of the degradation process is O(m*n), where m is the number of samples and n is the dimension.
[0105] (5) Figure 6 As shown in Table 2, the impact of algebra on accuracy during the backtracking process is illustrated. When the proposed algorithm is applied to the fault diagnosis of reciprocating compressor valves, the required diagnosis time is significantly reduced compared to machine learning algorithms, while maintaining high accuracy.
[0106] Table 2 Comparison of diagnostic results of different algorithms
[0107]
[0108] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the invention. Therefore, equivalent variations made according to the claims of the present invention are still within the scope of the invention. The present invention is a rapid fault diagnosis method for reciprocating compressors based on sparse grids and immune mechanisms. First, the collected signals are subjected to t-SNE dimensionality reduction visualization processing in the constructed cell grid space, optimally converting them into two-dimensional spatial data to reduce the complexity of the immune space. Simultaneously, a novel cell generation strategy (CG) is proposed to remove duplicate cells, avoiding the dimensionality explosion problem and reducing spatiotemporal complexity. Second, cell encoding function (CE) is assigned, and cell affinity is calculated using Manhattan distance, simplifying existing affinity calculation methods. Third, a continuous learning method based on a backtracking strategy (BC) is constructed to reduce repeated learning. Finally, a model adaptive adjustment method based on a single-layer feedforward neural network is used to achieve sample classification and clustering, improving the accuracy and efficiency of fault diagnosis classification.
[0109] The embodiments of the present invention have been described above, but the present invention is not limited to the above embodiments. Various changes can be made according to the purpose of the invention. Any changes, modifications, substitutions, combinations or simplifications made based on the spirit and principle of the technical solution of the present invention shall be equivalent substitutions. As long as they meet the purpose of the invention and do not deviate from the technical principle and inventive concept of the present invention, they shall fall within the protection scope of the present invention.
Claims
1. A rapid fault diagnosis method for reciprocating compressors based on sparse mesh and immune mechanism, characterized in that, include: S1. The cell grid space constructed from the original signals of the reciprocating compressor is subjected to t-SNE dimensionality reduction visualization processing to convert it into two-dimensional spatial data, which enhances the difference of heterogeneous samples and reduces the complexity of the immune space. A new cell generation strategy CG is proposed to delete duplicate cells, avoid the dimensionality explosion problem, and reduce the temporal and spatial complexity. S2. By endowing cells with encoding function CE, cell affinity is calculated through Manhattan distance, simplifying existing affinity calculation methods; secondly, a continuous learning method based on backtracking strategy BC is constructed to reduce repetitive learning. S3. An adaptive adjustment method for models based on single-layer feedforward neural networks is used to classify and cluster samples and diagnose the faults of reciprocating compressors. The step S1, which involves constructing a super-rectangular cell space and implementing a cell generation strategy (CG) for the original signal from the reciprocating compressor, specifically includes: S1-1. Using the proposed cell strategy, constant-size hyperrectangular cells are formed in the sample space. The cell size and position are determined by the cell generation and coordinates. The generated cells are used for subsequent encoding. S1-2, Antigen Sample: Antigen coordinates , , ; S1-3. Determine the location of the antigen in the cell: ; S1-4 Gaussian maximum integer function; S1-5. Describe cell coordinates using antigens: .
2. The rapid fault diagnosis method for reciprocating compressors based on sparse grids and immune mechanisms according to claim 1, characterized in that, The specific steps of cell-encoding CE in step S2 are as follows: S2-1-1, Encoding the cell; S2-1-2. Empty cells are recorded as 0 and do not participate in the calculation; only their positions are recorded. S2-1-3. Establish sparse matrices or tensors to obtain antigen information and record known cell tags; S2-1-4. Merge two cells with the same encoding into one cell. The new cell will overwrite the information of the old cell. S2-1-5. Retain sufficient information about outlier samples through reasonable algebraic settings; S2-1-6. When there is sufficient storage space or the dimensionality of the antigen is not high, tensor indexing is used to reduce time complexity.
3. The rapid fault diagnosis method for reciprocating compressors based on sparse grids and immune mechanisms according to claim 1, characterized in that, The Manhattan distance weight calculation step in step S2 is as follows: S2-2-1. To distinguish the contribution of each feature of an antigen to classification, the weight of each feature is calculated before calculating affinity; the purpose of increasing the weight is to maximize the affinity value of similar cells and minimize the affinity value of dissimilar cells; described as follows: ； In the formula: m and a Representing different cells, j and k Represents cell type; S2-2-2, A single-layer feedforward network training model was constructed, and the weights were updated using gradient descent with the squared loss function: ； In the formula: α Represents the learning rate. t This represents the number of iterations.
4. The rapid fault diagnosis method for reciprocating compressors based on sparse grids and immune mechanisms according to claim 1, characterized in that, The steps of the continuous learning method based on backtracking strategy BC in step S2 are as follows: S2-3-1, Basic Testing Process: Principles for Determining the Similarity of Unknown Antigen Types: ； In the formula: w i Represents feature weights, i . jk Representing different types of cells; S2-2-2, To prevent infinite affinity, calculations are performed in actual programming. aff -1; Therefore, the basic testing process is to calculate the weighted Manhattan distance between the current cell and the new cell activated by the unknown antigen; if the calculation result is zero, the new antigen must belong to the current cell; otherwise, the known antigen will be classified as the same as the cell with the highest affinity. S2-2-3. Cell generation will degenerate before the number of unknown classes matches the number of target classes. If the affinity between cells is the same, cells of known classes have a higher priority than cells of unknown classes, and cells activated first have a higher priority than cells activated later. S2-2-4. Backtracking strategy BC is a bottom-up approach that can only degenerate cells from high passage number to low passage number. The time complexity of the degradation process is O(m*n), where m is the number of samples and n is the dimension.