A method and system for intelligent identification of industrial hazards based on pattern recognition
By synchronously acquiring and analyzing multidimensional operating status data of industrial equipment in both time and space, a high-dimensional initial feature vector is generated and converted into a low-dimensional decoupled feature vector using an adaptive dimensionality reduction matrix. Combined with a local normal baseline model, this solves the problem of difficulty in identifying early hidden dangers under frequent load fluctuations in existing technologies, and achieves high accuracy and robustness in hidden danger identification.
Patent Information
- Application Number
- CN202611262305.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-19
- Publication Date
- 2026-11-17
AI Technical Summary
Existing pattern recognition systems struggle to effectively identify early hazard signals in scenarios with frequent fluctuations in equipment load, leading to overlap between early hazard samples and normal samples, making it difficult to distinguish weak hazard signals.
By collecting multidimensional operating status data during equipment operation, spatiotemporal synchronous acquisition and angular domain analysis are performed to generate high-dimensional initial feature vectors. These vectors are then converted into low-dimensional decoupled feature vectors through an adaptive dimensionality reduction mapping matrix. Combined with a local normal baseline model, a hazard index is calculated to identify early-stage hazards.
It significantly improves the accuracy and robustness of intelligent identification of industrial hazards, effectively separating the features dominated by load fluctuations from the weak features sensitive to hazards, and avoiding the problem of hazard features being submerged in load fluctuations by traditional fixed dimensionality reduction methods.
Smart Images

Figure CN123115754A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of pattern recognition technology, and more specifically, to a method and system for intelligent identification of industrial hazards based on pattern recognition. Background Technology
[0002] In industrial production, to achieve intelligent identification of potential hazards in complex equipment such as large reciprocating compressors, multi-dimensional sensor data, such as vibration acceleration and dynamic pressure, is typically collected during equipment operation and processed using pattern recognition technology. Because the raw data exhibits high-dimensionality and nonlinear characteristics, existing pattern recognition systems often first reduce the dimensionality of high-dimensional feature vectors using feature dimensionality reduction mapping matrices such as principal component analysis, and then perform distance matching with known hazard patterns in the low-dimensional space to output the identification result. This method can effectively distinguish between normal conditions and typical hazards when the equipment is running stably. However, in the actual operating scenario of reciprocating compressors experiencing frequent load fluctuations due to process adjustments, the feature dimensionality reduction mapping stage reveals significant limitations. In this scenario, the overall amplitude of vibration and pressure during normal operation changes drastically with the load, while early fatigue of the valve spring only causes a slight change in the pressure pulsation waveform within a very narrow range near the valve opening moment. Because fixed dimensionality reduction mapping matrices typically select principal components with the goal of maximizing global variance, when processing this type of data, the dimension with the largest variance in the overall amplitude fluctuation direction is assigned the highest projection weight. Meanwhile, the dimension bearing the minute characteristics of spring fatigue has a very low variance percentage, and its corresponding mapping coefficient is close to zero. This results in the minute pressure pulsation offsets representing slight degradation of spring stiffness having almost no component in the low-dimensional space, completely submerged by the projection direction dominated by load fluctuations. Early potential hazard samples degenerate into points overlapping with normal samples under wide operating conditions in the low-dimensional space. Existing pattern matching mechanisms struggle to distinguish weak hazard signals from these overlapping low-dimensional features, thus misjudging the degraded state as normal for a long time. Summary of the Invention
[0003] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application proposes an intelligent industrial hazard identification method and system based on pattern recognition, aiming to solve the problem in the prior art where, under scenarios with frequent fluctuations in equipment load, a fixed dimensionality reduction mapping matrix is difficult to effectively identify early hazard signals, leading to overlap between early hazard samples and normal samples, thus making it difficult to distinguish weak hazard signals.
[0004] In a first aspect, embodiments of this application provide a method for intelligent identification of industrial hazards based on pattern recognition, including:
[0005] Multidimensional operating status data is collected during the operation of the equipment, and the multidimensional operating status data is spatiotemporally acquired and angular domain analyzed to determine local sensitive data windows and global background data windows. The multidimensional operating status data includes at least cylinder dynamic pressure signal, valve cover vibration acceleration signal and crankshaft key phase pulse signal.
[0006] Global background features are extracted within the global background data window, and local sensitive features are extracted within the local sensitive data window. The extracted global background features and local sensitive features are concatenated and standardized to generate a high-dimensional initial feature vector.
[0007] A global background feature subset is extracted from the high-dimensional initial feature vector, and the distance between the global background feature subset and each historical anchor point in the offline pre-constructed load state baseline manifold is calculated to generate load background labels.
[0008] The high-dimensional initial feature vector is multiplied by a preset adaptive dimensionality reduction mapping matrix to obtain a low-dimensional decoupled feature vector, wherein the preset adaptive dimensionality reduction mapping matrix is obtained based on the load background label;
[0009] Based on the load background label, the corresponding local normal benchmark model is called from the pre-stored pattern matching model library, and the low-dimensional decoupled feature vector is input into the local normal benchmark model to obtain the hidden danger index.
[0010] The hazard index is compared with the judgment threshold of the local normal benchmark model to obtain the judgment result.
[0011] According to some embodiments of this application, the step of performing spatiotemporal synchronous acquisition and angular domain analysis on the multidimensional operating status data to determine the local sensitive data window and the global background data window includes:
[0012] Using the crankshaft key phase pulse signal as a synchronous trigger reference, the cylinder dynamic pressure signal and the valve cover vibration acceleration signal, which are sampled at equal time intervals, are converted into angular domain data based on equal crankshaft rotation angle sampling using an equal angle resampling algorithm.
[0013] Obtain the equipment design parameters and calculate the theoretical opening angle and theoretical closing angle of each air valve in the equipment based on the equipment design parameters;
[0014] The local sensitive data window is defined by expanding a predetermined angle range centered on the theoretical opening angle and the theoretical closing angle.
[0015] The global background data window is defined based on the complete rotation period of the theoretical opening angle and the theoretical closing angle.
[0016] According to some embodiments of this application, the step of extracting global background features within the global background data window and simultaneously extracting local sensitive features within the local sensitive data window includes:
[0017] Based on the global background data window, calculate the maximum value, minimum value, average indicated pressure, pressure ratio and volumetric efficiency of the cylinder dynamic pressure signal in a complete cycle, and at the same time calculate the root mean square value, peak value and total vibration energy of the valve cover vibration acceleration signal in the whole cycle to obtain global background characteristics.
[0018] Multi-layer wavelet packet decomposition is performed on the valve cover vibration acceleration signal within the local sensitive data window to divide the valve cover vibration acceleration signal into multiple high-frequency sub-bands. The energy proportion and energy entropy of the high-frequency sub-bands are calculated to obtain the local sensitive features.
[0019] According to some embodiments of this application, the preset adaptive dimensionality reduction mapping matrix is obtained through the following steps:
[0020] Based on the load background label, extract the normal sample set of the same load neighborhood and the pre-stored early fatigue prior sample set of the air valve spring from the historical database.
[0021] Based on the normal sample set and the pre-stored a priori sample set of early fatigue of the gas valve spring, the load fluctuation dominant subspace and the hidden danger sensitive subspace are constructed respectively.
[0022] The basis vectors of the hazard-sensitive subspace are orthogonalized to obtain the orthogonalized hazard-sensitive basis vectors.
[0023] The vectors of the load fluctuation dominant subspace are concatenated with the orthogonalized hidden danger sensitive basis vectors to generate an adaptive dimensionality reduction mapping matrix.
[0024] According to some embodiments of this application, the step of constructing a load fluctuation-dominant subspace and a hidden danger-sensitive subspace based on the normal sample set and the pre-stored early fatigue prior sample set of the gas valve spring, respectively, includes:
[0025] Obtain the global background features of the normal sample set and the local sensitive features of the pre-stored early fatigue prior sample set of the gas valve spring;
[0026] The covariance matrix is calculated based on the global background features of the normal sample set, and eigenvalue decomposition is performed. The eigenvector corresponding to the largest eigenvalue is extracted, and the load fluctuation dominant subspace is constructed.
[0027] The local covariance matrix is calculated based on the local sensitivity features of the pre-stored early fatigue prior sample set of the gas valve spring, and the main feature vector sensitive to spring degradation is extracted to construct the hidden danger sensitive subspace.
[0028] According to some embodiments of this application, the step of performing matrix multiplication on the high-dimensional initial feature vector and a preset adaptive dimensionality reduction mapping matrix to obtain a low-dimensional decoupled feature vector includes:
[0029] The high-dimensional initial feature vector is multiplied by the load-dominant basis vector in the preset adaptive dimensionality reduction mapping matrix to obtain the first low-dimensional feature vector, wherein the first low-dimensional feature vector represents the current macroscopic load migration status of the device.
[0030] The high-dimensional initial feature vector is multiplied by the hidden danger sensitive basis vector in the preset adaptive dimension reduction mapping matrix to obtain the second low-dimensional feature vector, wherein the second low-dimensional feature vector characterizes the degree of microscopic degradation of the local valve.
[0031] The first low-dimensional feature vector and the second low-dimensional feature vector are oriented and projected to reconstruct the low-dimensional decoupled feature space, and the low-dimensional decoupled feature vector is output.
[0032] According to some embodiments of this application, the step of inputting the low-dimensional decoupled feature vector into the local normal baseline model to obtain the hidden danger index includes:
[0033] A hypersphere is defined for the local normal baseline model, wherein the center and radius of the hypersphere characterize the normal data distribution under a specific load;
[0034] The low-dimensional decoupling feature vector is input into the local normal benchmark model, and the Euclidean distance of the low-dimensional decoupling feature vector from the center of the hypersphere in the low-dimensional decoupling feature space is calculated.
[0035] The hazard index is obtained based on the Euclidean distance, wherein the hazard index is positively correlated with the Euclidean distance.
[0036] According to some embodiments of this application, the step of calculating the distance between the global background feature subset and each historical anchor point in the offline pre-built load state baseline manifold, and generating load background labels, includes:
[0037] The historical fault-free data collected when the device undergoes various load adjustments under normal conditions is obtained. A subset of historical background features of the historical fault-free data is extracted and mapped to a low-dimensional embedding space through a manifold learning algorithm, so as to construct a load state reference manifold containing each historical anchor point and its corresponding embedding coordinates in the low-dimensional embedding space in an offline state.
[0038] The global background feature subset is mapped to the low-dimensional embedding space. The embedding space distance between the mapped feature points and each historical anchor point in the load state baseline manifold is calculated. The K-nearest neighbor search algorithm is used to find the nearest historical normal load samples and obtain the embedding space distance value of the nearest neighbor samples.
[0039] Obtain the embedding coordinates of each of the nearest neighbor samples in the low-dimensional embedding space, and perform a weighted summation of the embedding coordinates based on the reciprocal of the distance value in the embedding space as a weight to obtain the current embedding coordinate position of the global background feature subset in the low-dimensional embedding space;
[0040] Based on the mapping relationship between each embedded coordinate position and the load level value pre-calibrated in the load state reference manifold, the current embedded coordinate position is converted into a load background label.
[0041] According to some embodiments of this application, comparing the hazard index with the judgment threshold of the local normal benchmark model to obtain the discrimination result includes:
[0042] When the hidden danger index is less than or equal to the judgment threshold, the judgment result is that the current state is normal;
[0043] When the hidden danger index is greater than the judgment threshold, the offset vector is determined based on the difference between the low-dimensional decoupling feature vector and the standard center feature vector corresponding to the local normal benchmark model.
[0044] The offset vector is orthogonally decomposed into a preset sensitive projection direction to obtain the offset component along the sensitive projection direction;
[0045] When the offset component is less than or equal to the preset lower limit of sensitivity tolerance, the judgment result is normal operating condition drift;
[0046] When the offset component is greater than the lower limit of the sensitivity tolerance, the judgment result is that the valve spring has undergone early fatigue degradation.
[0047] Secondly, this application also discloses an intelligent industrial hazard identification system based on pattern recognition, comprising:
[0048] The acquisition module is used to acquire multi-dimensional operating status data during the operation of the equipment, and to perform spatiotemporal synchronous acquisition and angular domain analysis on the multi-dimensional operating status data to determine local sensitive data windows and global background data windows. The multi-dimensional operating status data includes at least cylinder dynamic pressure signal, valve cover vibration acceleration signal and crankshaft key phase pulse signal.
[0049] The extraction module is used to extract global background features from the global background data window and local sensitive features from the local sensitive data window. The extracted global background features and local sensitive features are concatenated and standardized to generate a high-dimensional initial feature vector.
[0050] The calculation module is used to extract a global background feature subset from the high-dimensional initial feature vector, calculate the distance between the global background feature subset and each historical anchor point in the offline pre-constructed load state baseline manifold, and generate load background labels.
[0051] The matrix multiplication module is used to perform matrix multiplication on the high-dimensional initial feature vector and the preset adaptive dimension reduction mapping matrix to obtain a low-dimensional decoupled feature vector, wherein the preset adaptive dimension reduction mapping matrix is obtained according to the load background label;
[0052] The input module is used to call the corresponding local normal benchmark model from the pre-stored pattern matching model library according to the load background label, and input the low-dimensional decoupled feature vector into the local normal benchmark model to obtain the hidden danger index;
[0053] The comparison module is used to compare the hidden danger index with the judgment threshold of the local normal benchmark model to obtain the judgment result.
[0054] According to the technical solution of this application embodiment, it has at least the following beneficial effects: By collecting multi-dimensional operating status data during equipment operation and performing spatiotemporal synchronous acquisition and angular domain analysis, local sensitive data windows and global background data windows are determined. Based on this, global background features are extracted within the global background data window, and local sensitive features are extracted within the local sensitive data window. The two are then concatenated and standardized to generate a high-dimensional initial feature vector. To address the problem in the prior art where a fixed dimensionality reduction mapping matrix is difficult to effectively identify early hidden danger signals in scenarios with frequent load fluctuations, this application extracts a subset of global background features from the high-dimensional initial feature vector, calculates the distance between this subset and each historical anchor point in the offline pre-constructed load state benchmark manifold, and generates a load background label. This load background label is used to dynamically adjust a preset adaptive dimensionality reduction mapping matrix, so that after performing matrix multiplication with the high-dimensional initial feature vector, a low-dimensional decoupled feature vector can be obtained. This adaptive dimensionality reduction mapping matrix can effectively separate the features dominated by load fluctuations and the weak features sensitive to hidden dangers according to the current load state, avoiding the problem of traditional fixed dimensionality reduction methods submerging hidden danger features in load fluctuations. Finally, based on the load background label, the corresponding local normal baseline model is invoked, and the low-dimensional decoupled feature vector is input into the model to obtain the hazard index, which is then compared with the judgment threshold to obtain the discrimination result. Through the above technical solution, this application can effectively solve the technical problem in the prior art that early hazard signals are difficult to identify in scenarios with frequent load fluctuations in equipment such as reciprocating compressors, and significantly improve the accuracy and robustness of intelligent identification of industrial hazards.
[0055] Additional aspects and advantages of this application 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 this application. Attached Figure Description
[0056] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.
[0057] Figure 1 A flowchart illustrating an embodiment of the intelligent identification method for industrial hazards based on pattern recognition provided in this application;
[0058] Figure 2 A flowchart illustrating the process of determining a local sensitive data window and a global background data window, provided in one embodiment of this application;
[0059] Figure 3 This is a schematic diagram illustrating the process of obtaining a preset adaptive dimensionality reduction mapping matrix according to one embodiment of this application;
[0060] Figure 4This is a schematic diagram of an industrial hazard intelligent identification system based on pattern recognition, provided as an embodiment of this application. Detailed Implementation
[0061] To make the objectives, technical methods, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0062] It should be noted that the meaning of "multiple" (or "more than") in the description of the embodiments of this application refers to two or more, and "greater than," "less than," "exceeding," etc. are understood to exclude the number itself, while "above," "below," "within," etc. are understood to include the number itself. If "first," "second," etc. are used in the description, they are only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.
[0063] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: the existence of a alone, the existence of b alone, the existence of c alone, the simultaneous existence of a and b, the simultaneous existence of a and c, the simultaneous existence of b and c, or the simultaneous existence of a, b, and c, where a, b, and c can be single or multiple.
[0064] In the description of this application, unless otherwise expressly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.
[0065] The pattern recognition-based intelligent identification method for industrial hazards provided in this application can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms; the software can be an application implementing the pattern recognition-based intelligent identification method for industrial hazards, etc., but is not limited to the above forms.
[0066] The embodiments of this application can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can reside in local and remote computer storage media, including storage devices.
[0067] See Figure 1 , Figure 1 This is a flowchart illustrating a pattern recognition-based intelligent identification method for industrial hazards provided in one embodiment of this application. The pattern recognition-based intelligent identification method for industrial hazards provided in this embodiment includes, but is not limited to, steps S110 to S160, which are described in detail below.
[0068] Step S110: Collect multi-dimensional operating status data during equipment operation, and perform spatiotemporal synchronous acquisition and angular domain analysis on the multi-dimensional operating status data to determine the local sensitive data window and the global background data window. The multi-dimensional operating status data includes at least the cylinder dynamic pressure signal, the valve cover vibration acceleration signal and the crankshaft key phase pulse signal.
[0069] Step S120: Extract global background features in the global background data window and extract local sensitive features in the local sensitive data window. Concatenate the extracted global background features and local sensitive features and perform standardization to generate a high-dimensional initial feature vector.
[0070] Step S130: Extract the global background feature subset from the high-dimensional initial feature vector, calculate the distance between the global background feature subset and each historical anchor point in the offline pre-constructed load state baseline manifold, and generate load background labels;
[0071] Step S140: Perform matrix multiplication on the high-dimensional initial feature vector and the preset adaptive dimension reduction mapping matrix to obtain the low-dimensional decoupled feature vector, wherein the preset adaptive dimension reduction mapping matrix is obtained according to the load background label;
[0072] Step S150: Based on the load background label, call the corresponding local normal baseline model from the pre-stored pattern matching model library, input the low-dimensional decoupled feature vector into the local normal baseline model, and obtain the hidden danger index.
[0073] Step S160: Compare the hidden danger index with the judgment threshold of the local normal benchmark model to obtain the judgment result.
[0074] It should be noted that "multidimensional operating status data" refers to various sensor data collected during the operation of industrial equipment, such as cylinder dynamic pressure signals, valve cover vibration acceleration signals, and crankshaft keyway pulse signals. This data can comprehensively reflect the operating status of the equipment. "Spatiotemporal synchronous acquisition and angular domain analysis" refers to aligning these multidimensional data in time and space, and transforming them from the time domain to the angular domain related to crankshaft rotation angle, in order to better capture features related to the mechanical motion of the equipment. "Local sensitive data window" refers to the data interval defined in the angular domain data for sensitive areas where early potential problems may occur in specific components (such as valves). "Global background data window" refers to a broader data interval in the angular domain data covering a complete working cycle of the equipment, used to capture the overall operating status and load information of the equipment. "Global background features" refers to features extracted from the global background data window that reflect the overall operating status and load level of the equipment. "Local sensitive features" refers to features extracted from the local sensitive data window that are sensitive to early potential problems in specific components. "High-dimensional initial feature vector" is a comprehensive feature representation formed by concatenating global background features and local sensitive features. The "Load State Baseline Manifold" is constructed offline and serves as a low-dimensional representation of the normal operating mode of equipment under different load conditions. The "Load Background Label" is determined based on the position of the current equipment's global background features within the Load State Baseline Manifold, indicating the equipment's current load state. The "Preset Adaptive Dimensionality Reduction Mapping Matrix" is a dynamically adjusted dimension reduction matrix based on the load background label, aiming to map high-dimensional feature vectors to a low-dimensional decoupled feature space. The "Low-Dimensional Decoupled Feature Vector" is obtained after adaptive dimension reduction, effectively separating load fluctuations and potential hazard features. The "Local Normal Baseline Model" is a model established for the equipment's normal operating mode under specific load conditions, used to assess the degree of anomaly in the current operating state. The "Hazard Index" is a quantitative indicator measuring the degree to which the current equipment's operating state deviates from the normal baseline. The "Judgment Threshold" is a preset limit used to determine whether the equipment has potential hazards.
[0075] In one embodiment, it is first necessary to collect multi-dimensional operating status data during the operation of the equipment. This data can be acquired by various sensors: pressure sensors are used to collect dynamic pressure signals from the cylinders, acceleration sensors are used to collect vibration acceleration signals from the valve cover, and photoelectric encoders or magnetoelectric sensors are used to collect crankshaft key phase pulse signals. These sensors can be installed in corresponding positions on the equipment; the pressure sensor is installed inside the cylinder, the acceleration sensor is installed on the surface of the valve cover, and the crankshaft key phase pulse sensor is installed near the crankshaft. The raw data collected is usually based on time-interval sampling. To better analyze periodic events related to crankshaft rotation angle, this data needs to be spatiotemporally synchronized for acquisition and angular domain analysis. For example, the crankshaft key phase pulse signal can be used as a synchronous trigger reference; when the crankshaft rotates through a preset angle, data from other sensors are collected synchronously. Subsequently, interpolation or resampling algorithms can be used to convert the time-domain data into angular-domain data; for example, the time-interval sampled cylinder dynamic pressure signal and valve cover vibration acceleration signal can be converted into angular-domain data based on crankshaft rotation angle sampling. In angular domain data, local sensitive data windows and global background data windows can be determined based on the equipment's structure and operating principles. For example, local sensitive data windows can be defined around the opening and closing moments of valves, as the valve's motion at these moments is most sensitive to early fatigue. Global background data windows, on the other hand, can cover a complete crankshaft rotation cycle to capture the overall operating status of the equipment.
[0076] In one embodiment, global background features can reflect the overall operating status and load of the equipment. For example, the maximum, minimum, average indicated pressure, pressure ratio, and volumetric efficiency of the cylinder dynamic pressure signal within a complete cycle can be calculated, as well as the root mean square value, peak-to-peak value, and total vibration energy of the valve cover vibration acceleration signal throughout the cycle. These features can be obtained through statistical analysis or signal processing methods. Locally sensitive features focus on capturing weak signals of early potential problems. For example, time-frequency analysis, such as wavelet analysis or short-time Fourier transform, can be performed on the valve cover vibration acceleration signal within the locally sensitive data window to extract features such as energy, entropy, or peak values within a specific frequency range. Multi-level wavelet packet decomposition can be performed on the valve cover vibration acceleration signal within the locally sensitive data window to divide the signal into multiple high-frequency sub-bands, and the energy proportion and energy entropy of these high-frequency sub-bands can be calculated. The extracted global background features and locally sensitive features are then concatenated to form a high-dimensional initial feature vector. Subsequently, a global background feature subset is extracted from the high-dimensional initial feature vector, and the distance between this global background feature subset and each historical anchor point in the offline pre-constructed load state baseline manifold is calculated. The load state baseline manifold can be constructed by manifold learning (such as Isomap, LLE, etc.) on historical fault-free data collected when the equipment undergoes various load adjustments under normal conditions. This maps the historical background feature subset to a low-dimensional embedding space, and records each historical anchor point and its corresponding embedding coordinates in the low-dimensional embedding space. By calculating the distance between the current global background feature subset's position in the low-dimensional embedding space and the historical anchor points, and combining this with the K-nearest neighbor search algorithm, the nearest historical normal load samples can be found. Based on the reciprocal of the embedding space distance values of these nearest neighbor samples as weights, their embedding coordinates are weighted and summed to obtain the current embedding coordinate position of the current global background feature subset in the low-dimensional embedding space. Finally, according to the pre-defined mapping relationship between each embedding coordinate position and the load level value in the load state baseline manifold, the current embedding coordinate position is converted into a load background label.
[0077] In one embodiment, a matrix multiplication operation is performed between the high-dimensional initial feature vector and a preset adaptive dimensionality reduction mapping matrix to obtain a low-dimensional decoupled feature vector. This preset adaptive dimensionality reduction mapping matrix is dynamically generated based on the previously obtained load background labels. For example, based on the load background labels, normal sample sets from the same load neighborhood and pre-stored a priori sample sets of early fatigue of valve springs can be extracted from a historical database. Based on these sample sets, a load fluctuation dominant subspace and a hidden danger sensitive subspace are constructed respectively. The load fluctuation dominant subspace can be obtained by performing principal component analysis (PCA) on the global background features of the normal sample set and extracting the feature vectors corresponding to the largest eigenvalues. The hidden danger sensitive subspace can be constructed by calculating the local covariance matrix of the local sensitive features of the early fatigue of valve springs and extracting the principal eigenvectors sensitive to spring degradation. The basis vectors of the hidden danger sensitive subspace are orthogonalized, and then the vectors of the load fluctuation dominant subspace are concatenated with the orthogonalized hidden danger sensitive basis vectors to generate the adaptive dimensionality reduction mapping matrix. During matrix multiplication, the high-dimensional initial feature vector is multiplied by the load-dominant basis vector in the pre-defined adaptive dimensionality reduction mapping matrix to obtain the first low-dimensional feature vector, representing the current macroscopic load migration status of the equipment. Simultaneously, the high-dimensional initial feature vector is multiplied by the hidden danger-sensitive basis vector in the pre-defined adaptive dimensionality reduction mapping matrix to obtain the second low-dimensional feature vector, representing the local microscopic degradation degree of the valve. Finally, the first and second low-dimensional feature vectors are oriented and projected to reconstruct the low-dimensional decoupled feature space, outputting the low-dimensional decoupled feature vector. Based on the load background label, the corresponding local normal baseline model is called from the pre-stored pattern matching model library. The pattern matching model library stores models trained for normal operating modes of the equipment under different load conditions. For example, a local normal baseline model can be trained for each load level or load neighborhood. These models can employ One-Class Support Vector Machines (SVM), Gaussian Mixture Models (GMM), or hyperspherical models, etc. For example, for a hyperspherical model, a hypersphere can be defined, whose center and radius represent the normal data distribution under a specific load. The low-dimensional decoupling feature vector is input into the invoked local normal baseline model, and the Euclidean distance of the low-dimensional decoupling feature vector from the center of the hypersphere in the low-dimensional decoupling feature space is calculated. A hazard index is obtained based on this Euclidean distance, where the hazard index is positively correlated with the Euclidean distance; that is, the farther the distance from the center of the hypersphere, the higher the hazard index.
[0078] In one embodiment, the hazard index is compared with the judgment threshold of the local normal baseline model to obtain a discrimination result. The judgment threshold is preset based on historical data and expert experience. When the hazard index is less than or equal to the judgment threshold, the discrimination result is that the current state is normal. When the hazard index is greater than the judgment threshold, further analysis is required. At this time, the offset vector can be determined based on the difference between the low-dimensional decoupled feature vector and the standard central feature vector corresponding to the local normal baseline model. The offset vector is orthogonally decomposed into a preset sensitive projection direction to obtain the offset component along the sensitive projection direction. The sensitive projection direction can be predetermined by expert knowledge or data-driven methods; for example, the direction can be aligned with the direction of the hazard sensitive subspace. When the offset component is less than or equal to the preset lower limit of sensitivity tolerance, the discrimination result is normal operating condition drift, indicating that although the equipment state deviates from normal, it is not caused by a specific hazard. When the offset component is greater than the lower limit of sensitivity tolerance, the discrimination result is that the valve spring has undergone early fatigue degradation, thereby achieving accurate identification of specific hazards.
[0079] It should be noted that this application also introduces a local normal baseline model. The corresponding local normal baseline model is invoked based on the load background label, and the low-dimensional decoupled feature vector is input into the model to obtain the hazard index. This localized pattern matching strategy enables modeling of the normal data distribution under different load conditions, improving the accuracy of hazard identification. Finally, by comparing the hazard index with a judgment threshold and further orthogonally decomposing the offset vector towards the sensitive projection direction, accurate discrimination of normal operating condition drift and early fatigue degradation of valve springs is achieved. Compared with existing technologies, the method of this application can effectively address the challenges of industrial equipment under frequent load fluctuations, significantly improving the accuracy and robustness of early hazard identification, and providing more reliable technical support for predictive maintenance of industrial equipment.
[0080] See Figure 2 , Figure 2 This is a flowchart illustrating the process of determining a local sensitive data window and a global background data window, as provided in one embodiment of this application. This application further proposes the steps of performing spatiotemporal synchronous acquisition and angular domain analysis of multidimensional operational status data, determining the local sensitive data window and the global background data window, and generating a predicted pressure-displacement curve.
[0081] Step S210: Using the crankshaft key phase pulse signal as a synchronous trigger reference, the cylinder dynamic pressure signal and valve cover vibration acceleration signal based on time-equal interval sampling are converted into angular domain data based on crankshaft rotation angle equal interval sampling using an equal angle resampling algorithm.
[0082] Step S220: Obtain the equipment design parameters and calculate the theoretical opening angle and theoretical closing angle of each air valve in the equipment based on the equipment design parameters;
[0083] Step S230: Define a local sensitive data window by expanding a predetermined angle range centered on the theoretical opening angle and theoretical closing angle;
[0084] Step S240: Define the global background data window based on the complete rotation period of the theoretical opening angle and the theoretical closing angle.
[0085] Using the crankshaft key phase pulse signal as a synchronization trigger reference means that during data acquisition, a specific angular position of the crankshaft is used as the synchronization signal to ensure that the data collected by different sensors are precisely aligned in time. The equal-angle resampling algorithm can be understood as converting the original time-domain data into data evenly distributed across the crankshaft rotation domain through interpolation or resampling. Equipment design parameters refer to inherent parameters related to the equipment structure and operation, such as the geometry of the cylinder and the movement patterns of the valves. Based on these design parameters, the theoretical opening and closing angles of each valve in the equipment under normal operating conditions can be accurately calculated. Defining local sensitive data windows by expanding a predetermined angle range centered on the theoretical opening and closing angles means using the theoretical opening and closing angles of the valves as a reference, extending a preset, smaller angle range forward and backward to form one or more local data windows. These local sensitive data windows are specifically used to capture the dynamic response of the valves at the moment of opening or closing, which are often the areas where valve failures (such as spring fatigue) are most likely to manifest abnormalities. The setting of the predetermined angle range can be optimized based on experience or experiments to ensure sufficient coverage of the sensitive areas of the valve's critical actions. Furthermore, defining the global background data window based on the complete rotational cycle of the theoretical opening and closing angles means using one complete rotational cycle of the crankshaft (e.g., 0 degrees to 360 degrees) as the global background data window. This global background data window is used to capture the macroscopic operating status of the equipment throughout the entire working cycle, including background information such as load changes and overall vibration levels. By defining the global background data window, a comprehensive understanding of the overall operating characteristics of the equipment can be obtained, providing a foundation for subsequent load condition identification and feature decoupling.
[0086] It should be noted that the solution proposed in this application effectively solves the problem of speed fluctuations interfering with feature extraction in traditional time-domain analysis by transforming multidimensional operating status data from the time domain to the angular domain and precisely defining local sensitive data windows and global background data windows. Specifically, by using the crankshaft key phase pulse signal as a synchronous trigger reference and employing an equal-angle resampling algorithm, the acquired cylinder dynamic pressure signal and valve cover vibration acceleration signal are consistent in the crankshaft rotation angular domain, thereby eliminating the signal stretching or compression effect caused by changes in equipment speed and ensuring the stability and comparability of features. Based on this, by obtaining equipment design parameters and calculating the theoretical opening and closing angles of the valves, a precise physical reference is provided for subsequent data window definition. By expanding a predetermined angle range centered on these theoretical angles to define the local sensitive data windows, the system can focus on the critical moments of valve action and capture weak abnormal signals related to early fatigue degradation of the valves. At the same time, defining the global background data window with the complete rotation cycle ensures comprehensive monitoring of the overall operating status of the equipment and provides necessary background information for distinguishing between load fluctuations and potential degradation.
[0087] Specifically, the steps of extracting global background features within the global background data window and extracting local sensitive features within the local sensitive data window include:
[0088] The maximum, minimum, average indicated pressure, pressure ratio and volumetric efficiency of the cylinder dynamic pressure signal in a complete cycle are calculated based on the global background data window. At the same time, the root mean square value, peak-to-peak value and total vibration energy of the valve cover vibration acceleration signal in the entire cycle are calculated to obtain the global background characteristics.
[0089] Multi-level wavelet packet decomposition is performed on the valve cover vibration acceleration signal within the local sensitive data window, dividing the valve cover vibration acceleration signal into multiple high-frequency sub-bands. The energy proportion and energy entropy of the high-frequency sub-bands are calculated to obtain the local sensitive features.
[0090] The extraction of global background features aims to comprehensively reflect the macroscopic operating status and load conditions of the equipment. Specifically, for the cylinder dynamic pressure signal, its maximum, minimum, average indicated pressure, pressure ratio, and volumetric efficiency directly characterize the engine's combustion process, work capacity, and intake and exhaust efficiency. These parameters are closely related to the equipment's load conditions. Simultaneously, for the valve cover vibration acceleration signal, its root mean square value, peak-to-peak value, and total vibration energy over the entire cycle reflect the overall vibration level and energy distribution of the equipment. These features are crucial for assessing the overall health and load fluctuations of the equipment. Furthermore, the extraction of local sensitive features focuses on capturing early signs of localized, microscopic degradation in the equipment. Multi-level wavelet packet decomposition of the valve cover vibration acceleration signal within the local sensitive data window decomposes the original signal into different frequency sub-bands, enabling refined signal analysis in the time-frequency domain. By dividing the signal into multiple high-frequency sub-bands, the impact and transient components associated with early failures (such as valve spring fatigue) can be effectively highlighted. Subsequently, the energy proportion and energy entropy of these high-frequency sub-bands are calculated, which can quantify the energy distribution and signal complexity within a specific frequency range. These indicators are highly sensitive to minute structural changes and energy anomalies, and are key to identifying early potential problems.
[0091] It should be noted that the solution presented in this application achieves a comprehensive and detailed characterization of the equipment's operating status through the aforementioned feature extraction methods. The extraction of global background features enables the system to accurately perceive macroscopic load changes and overall operating trends, laying the foundation for subsequent load decoupling. Simultaneously, the extraction of locally sensitive features, particularly through multi-layer wavelet packet decomposition and high-frequency sub-band energy analysis, effectively amplifies and captures early, weak fault signals that are difficult to detect using traditional methods, such as early fatigue degradation of valve springs. This hierarchical, multi-scale feature extraction strategy allows the system to effectively distinguish between macroscopic load fluctuations and microscopic fault degradation in complex industrial operating environments, thereby avoiding misjudgments caused by load changes and improving the ability to identify early potential hazards.
[0092] See Figure 3 , Figure 3 This is a schematic diagram of a process for obtaining a preset adaptive dimensionality reduction mapping matrix according to an embodiment of this application.
[0093] Step S310: Based on the load background label, extract the normal sample set of the same load neighborhood and the pre-stored early fatigue prior sample set of the air valve spring from the historical database.
[0094] Step S320: Based on the normal sample set and the pre-stored a priori sample set of early fatigue of air valve springs, construct the load fluctuation dominant subspace and the hidden danger sensitive subspace respectively.
[0095] Step S330: Orthogonalize the basis vectors of the hazard-sensitive subspace to obtain the orthogonalized hazard-sensitive basis vectors;
[0096] Step S340: Concatenate the vectors of the load fluctuation dominant subspace with the orthogonalized hidden danger sensitive basis vectors to generate an adaptive dimensionality reduction mapping matrix.
[0097] Specifically, the load background label is obtained by calculating the distance between a subset of global background features and each historical anchor point in the offline pre-constructed load state baseline manifold, representing the current macroscopic operating load state of the equipment. Based on this load background label, a normal sample set matching the current load state can be accurately selected from the historical database. These samples represent the feature distribution of the equipment operating without failure under this load condition. Meanwhile, the pre-stored early fatigue prior sample set of valve springs contains typical feature patterns of early fatigue degradation of valve springs under different loads. This prior knowledge is crucial for identifying weak early potential hazard signals. The construction of the load fluctuation dominant subspace aims to capture the main direction of data changes under normal load fluctuations, while the construction of the hazard-sensitive subspace focuses on extracting weak but critical feature directions related to early fatigue degradation of valve springs. To ensure that these two subspaces can independently represent their respective physical meanings and avoid the obscuring of hazard features by load fluctuations, the basis vectors of the hazard-sensitive subspace need to be orthogonalized. Orthogonalization eliminates the influence of the load fluctuation-dominant subspace on the hazard-sensitive subspace, allowing hazard features to be clearly separated in the low-dimensional space. Finally, the orthogonally-processed hazard-sensitive basis vectors are concatenated with the vectors of the load fluctuation-dominant subspace to generate an adaptive dimensionality-reduced mapping matrix that simultaneously reflects load fluctuations and hazard-sensitive directions. This matrix can be adjusted according to the current load background label to adapt to different operating conditions.
[0098] It should be noted that this application can adaptively construct a dimension-reduced mapping matrix based on the real-time load status of the equipment, thereby effectively decoupling the macroscopic load fluctuation characteristics from the microscopic hazard sensitivity characteristics. This method significantly enhances the system's ability to identify weak hazard signals such as early fatigue of valve springs, reduces the false alarm rate caused by load changes, and improves the accuracy and robustness of hazard identification. Especially in the complex and variable operating environment of industrial equipment, the construction method of this adaptive dimension-reduced mapping matrix ensures that the hazard identification model always focuses on the true failure mode, rather than being limited by normal operating condition drift, thus providing a more reliable decision-making basis for predictive maintenance of equipment.
[0099] Specifically, based on the normal sample set and the pre-stored a priori sample set of early fatigue of the gas valve spring, the above-mentioned load fluctuation dominant subspace and hidden danger sensitive subspace are constructed respectively, including the following steps:
[0100] Obtain the global background features of the normal sample set and the local sensitive features of the pre-stored early fatigue prior sample set of the gas valve spring.
[0101] The covariance matrix is calculated based on the global background features of the normal sample set and eigenvalue decomposition is performed. The eigenvector corresponding to the largest eigenvalue is extracted to construct the load fluctuation dominant subspace.
[0102] The local covariance matrix is calculated based on the local sensitivity features of the pre-stored early fatigue prior sample set of gas valve springs, and the main feature vector sensitive to spring degradation is extracted to construct the hidden danger sensitive subspace.
[0103] Specifically, when constructing the load fluctuation-dominant subspace, the first step is to acquire a normal sample set collected during various load adjustments while the equipment is operating normally. The global background features extracted from this normal sample set aim to comprehensively characterize the macroscopic operating state of the equipment under different load conditions. Subsequently, the covariance matrix is calculated based on these global background features. This covariance matrix quantifies the variance and covariance relationships between different global background features, reflecting the distribution and trend of data in the feature space. By performing eigenvalue decomposition on this covariance matrix, a series of eigenvalues and corresponding eigenvectors can be obtained. The eigenvector corresponding to the largest eigenvalue represents the direction of the most significant data change, which is usually closely related to the macroscopic fluctuations of the equipment load. Therefore, extracting these eigenvectors corresponding to the largest eigenvalues allows the construction of the load fluctuation-dominant subspace. This subspace can effectively capture and characterize the main data patterns of the equipment under normal load changes. Furthermore, when constructing the hazard-sensitive subspace, it is necessary to acquire a pre-stored prior sample set of early fatigue of the valve springs. This sample set contains the local sensitive features of the valve springs when early fatigue degradation occurs. These locally sensitive features were carefully selected and extracted, demonstrating a high degree of sensitivity to the microscopic degradation of the valve spring. Based on these locally sensitive features, a local covariance matrix was calculated. This local covariance matrix focuses on analyzing local feature changes related to early fatigue degradation of the valve spring. By analyzing this local covariance matrix, principal eigenvectors highly sensitive to spring degradation can be extracted.
[0104] This application's solution constructs a load fluctuation-dominant subspace and a hidden danger-sensitive subspace by utilizing the global background features of the normal sample set and the local sensitive features of the pre-stored early fatigue prior sample set of valve springs, respectively. This effectively decouples load change information and hidden danger degradation information during equipment operation. The load fluctuation-dominant subspace accurately captures the main data change direction of the equipment under normal load variations, thereby separating the impact of macroscopic load migration from fault diagnosis. Simultaneously, the hidden danger-sensitive subspace focuses on identifying and characterizing the microscopic feature changes unique to early fatigue degradation of valve springs, ensuring accurate detection of potential early hidden dangers even under large load fluctuations. This dual-subspace construction strategy allows the subsequent adaptive dimensionality reduction mapping matrix to more accurately project the high-dimensional initial feature vector onto the low-dimensional decoupled feature space, thus achieving effective separation of load and hidden danger information.
[0105] In response, this application further proposes a step of performing matrix multiplication on the aforementioned high-dimensional initial feature vector and a preset adaptive dimensionality reduction mapping matrix to obtain a low-dimensional decoupled feature vector, including:
[0106] The first low-dimensional feature vector is obtained by performing an inner product operation between the high-dimensional initial feature vector and the load-dominant basis vector in the preset adaptive dimensionality reduction mapping matrix. The first low-dimensional feature vector represents the current macroscopic load migration status of the equipment.
[0107] The second low-dimensional feature vector is obtained by performing an inner product operation between the high-dimensional initial feature vector and the hidden danger sensitive basis vector in the preset adaptive dimension reduction mapping matrix. The second low-dimensional feature vector represents the degree of micro-degradation of the local valve.
[0108] The first and second low-dimensional feature vectors are oriented and projected to reconstruct the low-dimensional decoupled feature space, and the low-dimensional decoupled feature vectors are output.
[0109] Specifically, the load-dominant basis vectors refer to the set of basis vectors in the preset adaptive dimensionality reduction mapping matrix that best reflects the load change trend and state of the equipment. By performing an inner product operation between the high-dimensional initial feature vector and these load-dominant basis vectors, the components related to macroscopic load migration in the original high-dimensional data can be effectively extracted to form a first low-dimensional feature vector. The numerical change of this first low-dimensional feature vector will directly reflect the macroscopic operating state of the equipment under different load conditions. The hidden danger-sensitive basis vectors refer to the set of basis vectors in the preset adaptive dimensionality reduction mapping matrix that are most sensitive to microscopic hidden dangers such as early fatigue degradation of key components like valve springs. Performing an inner product operation between the high-dimensional initial feature vector and these hidden danger-sensitive basis vectors can extract feature components closely related to the degree of local microscopic degradation of the valve from the complex operating data, generating a second low-dimensional feature vector. The numerical change of this second low-dimensional feature vector will directly indicate whether there are any abnormalities or degradation trends in the local area of the valve. In practical applications, directional projection of the first and second low-dimensional feature vectors refers to ensuring, through specific mathematical transformations, that these two feature vectors possess maximum independence and discriminability in their respective dimensions within the decoupled feature space. For example, orthogonal projection or non-orthogonal projection methods with clear physical meaning can be used to map them to a new low-dimensional space, thereby reconstructing a low-dimensional decoupled feature space that clearly distinguishes between the impact of macroscopic loads and the impact of microscopic hazards. Ultimately, the output low-dimensional decoupled feature vector is the result of this directional projection processing, containing two key, decoupled pieces of information regarding load and hazards in the equipment's operating state.
[0110] In one embodiment, this application achieves a high degree of decoupling between macroscopic load migration and microscopic hazard degradation during equipment operation. This decoupling frees the hazard identification process from the complexity caused by load changes, significantly improving the identification accuracy and robustness of hazards such as early fatigue degradation. Furthermore, since the first low-dimensional feature vector and the second low-dimensional feature vector correspond to macroscopic load and microscopic degradation respectively, this solution also enhances the interpretability of hazard diagnosis, enabling engineers to more intuitively understand the causes of anomalies, thereby providing strong support for accurate maintenance and fault prediction.
[0111] Specifically, the steps described above for inputting the low-dimensional decoupled feature vector into the local normal baseline model to obtain the hazard index include:
[0112] A hypersphere is defined for the local normal baseline model, where the center and radius of the hypersphere characterize the normal data distribution under a specific load;
[0113] Input the low-dimensional decoupled feature vector into the local normal baseline model and calculate the Euclidean distance of the low-dimensional decoupled feature vector from the center of the hypersphere in the low-dimensional decoupled feature space.
[0114] The hazard index is obtained based on the Euclidean distance, and the hazard index is positively correlated with the Euclidean distance.
[0115] The hypersphere can be understood as the geometric boundary in a low-dimensional decoupled feature space that defines the range of normal operating data for a device under a specific load condition. The center of the hypersphere is typically calculated by clustering or averaging the low-dimensional decoupled feature vectors of a large number of normal operating samples, representing the typical state of normal data under that specific load. The radius of the hypersphere is determined based on the dispersion of the normal data or a preset confidence level, encompassing the vast majority of normal data points and thus characterizing the range of normal data distribution under a specific load. Furthermore, when the low-dimensional decoupled feature vector is input into the local normal baseline model, the system calculates the Euclidean distance between the feature vector and the center of the hypersphere.
[0116] In one embodiment, the present application's solution defines a hypersphere in a local normal baseline model, using its center and radius to characterize the normal data distribution under a specific load, thereby effectively quantifying the deviation of the current equipment operating state from the normal state. When the equipment's operating state deviates from normal, the position of its corresponding low-dimensional decoupling feature vector in the low-dimensional decoupling feature space will be far from the center of the hypersphere, resulting in an increase in the calculated Euclidean distance. Therefore, by establishing a positive correlation between the Euclidean distance and the hazard index, the hazard index can intuitively reflect the potential degree of abnormality or degradation of the equipment.
[0117] In some embodiments of this application, the above method includes calculating the distance between a subset of global background features and each historical anchor point in an offline pre-built load state baseline manifold, and generating load background labels. Specifically, calculating the distance between a subset of global background features and each historical anchor point in an offline pre-built load state baseline manifold, and generating load background labels, includes the following steps:
[0118] In one embodiment, historical fault-free data collected during various load adjustments under normal operating conditions is acquired. A subset of historical background features from this data is extracted, and a manifold learning algorithm is used to map this subset to a low-dimensional embedding space. This allows for the construction of a load state baseline manifold offline, containing historical anchor points and their corresponding embedding coordinates in the low-dimensional embedding space. Historical fault-free data refers to a dataset accumulated by the equipment under different load conditions during normal operation, free from any faults or potential problems. Extracting the historical background feature subset aims to capture the macroscopic operating characteristics of the equipment under normal load variations. Manifold learning algorithms, such as Locally Linear Embedding (LLE), Isomap, or t-SNE, are used to reduce the dimensionality of the high-dimensional historical background feature subset and map it to a low-dimensional embedding space, thereby revealing the inherent nonlinear structure of the data. The global background feature subset is mapped to the low-dimensional embedding space, and the embedding space distance between the mapped feature points and each historical anchor point in the load state baseline manifold is calculated. The K-nearest neighbor search algorithm is used to find the nearest historical normal load samples, obtaining the embedding space distance values of the nearest neighbor samples. Specifically, the currently collected subset of global background features is mapped into the same low-dimensional embedding space using the same manifold learning algorithm as when constructing the load state baseline manifold, forming a current feature point.
[0119] In one embodiment, the Euclidean distance or other suitable distance metric between the current feature point and all historical anchor points on the baseline manifold is calculated to quantify their similarity. A K-nearest neighbor search algorithm is applied to identify the K closest historical normal load samples to the current feature point; the embedding space distance values of these samples are used for subsequent load background label generation. The embedding coordinates of each nearest sample in the low-dimensional embedding space are obtained. Based on the reciprocal of the embedding space distance value as a weight, the embedding coordinates are weighted and summed to obtain the current embedding coordinate position of the global background feature subset in the low-dimensional embedding space. The closer the sample, the greater its weight and the stronger its influence on the current embedding coordinate position, thus enabling the current embedding coordinate position to more accurately reflect the current load status of the device. According to the pre-defined mapping relationship between each embedding coordinate position and the load level value in the load status baseline manifold, the current embedding coordinate position is converted into a load background label. When constructing the load status baseline manifold, different regions or anchor points on the manifold are usually pre-labeled with their corresponding load level values.
[0120] In one embodiment, the above technical solution enables refined identification and quantification of equipment load status. This load background label generation mechanism based on manifold learning and K-nearest neighbor search effectively addresses the problems of large load fluctuations and complex and changing operating conditions during industrial equipment operation, providing accurate load background information for subsequent hazard identification and diagnosis. Therefore, it ensures that under different load conditions, the hazard identification model can call the benchmark model that best matches the current load status, thereby improving the accuracy and reliability of hazard identification and reducing the risk of false alarms and missed alarms.
[0121] The above comparison of the hazard index with the judgment threshold of the local normal baseline model yields the discrimination result, which specifically includes:
[0122] When the hazard index is less than or equal to the judgment threshold, the judgment result is determined to be that the current state is normal;
[0123] When the hazard index is greater than the judgment threshold, the offset vector is determined based on the difference between the low-dimensional decoupling feature vector and the standard central feature vector corresponding to the local normal benchmark model.
[0124] The offset vector is orthogonally decomposed into the preset sensitive projection direction to obtain the offset component along the sensitive projection direction.
[0125] When the offset component is less than or equal to the preset lower limit of sensitivity tolerance, the judgment result is determined to be normal operating condition drift;
[0126] When the offset component is greater than the lower limit of the sensitivity tolerance, the discrimination result is determined to be early fatigue degradation of the valve spring.
[0127] Specifically, the hazard index is a quantitative indicator that measures the degree to which the current operating state of equipment deviates from the normal baseline. The larger the value, the greater the degree of deviation from the normal state. The judgment threshold is a pre-set critical value used to distinguish between normal and abnormal states. When the hazard index exceeds this threshold, it indicates that the equipment may be abnormal. When the hazard index is greater than the judgment threshold, an offset vector needs to be introduced to further distinguish the nature of the abnormality. The offset vector refers to the difference between the current low-dimensional decoupled feature vector and the standard normal state center feature vector represented by the local normal baseline model. It characterizes the specific direction and magnitude of the current abnormal state in the feature space. The preset sensitive projection direction is determined based on historical data and expert experience and is used to identify the main change direction of specific hazards (such as early fatigue degradation of valve springs) in the feature space. By orthogonally decomposing the offset vector towards this sensitive projection direction, the offset component along this direction can be obtained. This component can effectively highlight the feature changes related to specific hazards. The lower limit of sensitivity tolerance is a threshold set for the offset component in the sensitive projection direction. This threshold is used to distinguish between the offset caused by fluctuations in normal operating conditions and the offset caused by early degradation of specific components (such as valve springs).
[0128] In one embodiment, the present application's solution addresses the lack of specificity in the discrimination results of the basic solution by introducing a more refined analysis of abnormal states. Specifically, when the hazard index exceeds the judgment threshold, indicating an abnormality in the equipment, instead of simply outputting an "abnormal" result, it quantifies the deviation between the current state and the normal baseline by calculating an offset vector. Furthermore, by orthogonally decomposing this offset vector towards a preset sensitive projection direction, it is possible to effectively separate feature changes related to specific hazards (such as early fatigue degradation of valve springs) from other types of operating condition drift. This is because different types of abnormalities typically evolve along different directions in the feature space. Through the preset sensitive projection direction, the feature changes most relevant to valve spring degradation can be focused on. Finally, by comparing the offset component along the sensitive projection direction with the lower limit of the sensitive tolerance, it is possible to accurately distinguish between normal operating condition drift and early fatigue degradation of the valve spring. This hierarchical discrimination mechanism enables the system to provide more instructive diagnostic information.
[0129] See Figure 4 , Figure 4 This is a schematic diagram of a pattern recognition-based intelligent industrial hazard identification system according to an embodiment of this application. The pattern recognition-based intelligent industrial hazard identification system 400 includes:
[0130] The acquisition module 410 is used to acquire multi-dimensional operating status data during the operation of the equipment, and to perform spatiotemporal synchronous acquisition and angular domain analysis on the multi-dimensional operating status data to determine the local sensitive data window and the global background data window. The multi-dimensional operating status data includes at least the cylinder dynamic pressure signal, the valve cover vibration acceleration signal and the crankshaft key phase pulse signal.
[0131] The extraction module 420 is used to extract global background features from the global background data window and local sensitive features from the local sensitive data window. The extracted global background features and local sensitive features are concatenated and standardized to generate a high-dimensional initial feature vector.
[0132] The calculation module 430 is used to extract a global background feature subset from the high-dimensional initial feature vector, calculate the distance between the global background feature subset and each historical anchor point in the offline pre-built load state baseline manifold, and generate load background labels.
[0133] The matrix multiplication module 440 is used to perform matrix multiplication on the high-dimensional initial feature vector and the preset adaptive dimension reduction mapping matrix to obtain the low-dimensional decoupled feature vector. The preset adaptive dimension reduction mapping matrix is obtained according to the load background label.
[0134] The input module 450 is used to call the corresponding local normal baseline model from the pre-stored pattern matching model library according to the load background label, input the low-dimensional decoupled feature vector into the local normal baseline model, and obtain the hidden danger index.
[0135] The comparison module 460 is used to compare the hazard index with the judgment threshold of the local normal benchmark model to obtain the judgment result.
[0136] It should be noted that the information interaction and execution process between the above modules are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.
[0137] Those skilled in the art will understand that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically include computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.
[0138] The above provides a detailed description of the preferred embodiments of this application. However, this application is not limited to the above-described embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined in this application.
Claims
1. A method for intelligent identification of industrial hazards based on pattern recognition, characterized in that, include: Multidimensional operating status data is collected during the operation of the equipment, and the multidimensional operating status data is spatiotemporally collected and angular domain analyzed to determine local sensitive data windows and global background data windows. The multidimensional operating status data includes at least cylinder dynamic pressure signal, valve cover vibration acceleration signal and crankshaft key phase pulse signal. Global background features are extracted within the global background data window, and local sensitive features are extracted within the local sensitive data window. The extracted global background features and local sensitive features are concatenated and standardized to generate a high-dimensional initial feature vector. A global background feature subset is extracted from the high-dimensional initial feature vector, and the distance between the global background feature subset and each historical anchor point in the offline pre-constructed load state baseline manifold is calculated to generate load background labels. The high-dimensional initial feature vector is multiplied by a preset adaptive dimensionality reduction mapping matrix to obtain a low-dimensional decoupled feature vector, wherein the preset adaptive dimensionality reduction mapping matrix is obtained based on the load background label; Based on the load background label, the corresponding local normal benchmark model is called from the pre-stored pattern matching model library, and the low-dimensional decoupled feature vector is input into the local normal benchmark model to obtain the hidden danger index. The hazard index is compared with the judgment threshold of the local normal benchmark model to obtain the judgment result.
2. The method according to claim 1, characterized in that, The process of spatiotemporally synchronizing the acquisition and angular domain analysis of the multidimensional operational status data to determine the local sensitive data window and the global background data window includes: Using the crankshaft key phase pulse signal as a synchronous trigger reference, the cylinder dynamic pressure signal and the valve cover vibration acceleration signal, which are sampled at equal time intervals, are converted into angular domain data based on equal crankshaft rotation angle sampling using an equal angle resampling algorithm. Obtain the equipment design parameters and calculate the theoretical opening angle and theoretical closing angle of each air valve in the equipment based on the equipment design parameters; The local sensitive data window is defined by expanding a predetermined angle range centered on the theoretical opening angle and the theoretical closing angle. The global background data window is defined based on the complete rotation period of the theoretical opening angle and the theoretical closing angle.
3. The method according to claim 1, characterized in that, The step of extracting global background features within the global background data window and simultaneously extracting local sensitive features within the local sensitive data window includes: Based on the global background data window, calculate the maximum value, minimum value, average indicated pressure, pressure ratio and volumetric efficiency of the cylinder dynamic pressure signal in a complete cycle, and at the same time calculate the root mean square value, peak value and total vibration energy of the valve cover vibration acceleration signal in the whole cycle to obtain global background characteristics. Multi-layer wavelet packet decomposition is performed on the valve cover vibration acceleration signal within the local sensitive data window to divide the valve cover vibration acceleration signal into multiple high-frequency sub-bands. The energy proportion and energy entropy of the high-frequency sub-bands are calculated to obtain the local sensitive features.
4. The method according to claim 1, characterized in that, The preset adaptive dimension reduction mapping matrix is obtained through the following steps: Based on the load background label, extract the normal sample set of the same load neighborhood and the pre-stored early fatigue prior sample set of the air valve spring from the historical database. Based on the normal sample set and the pre-stored a priori sample set of early fatigue of the gas valve spring, the load fluctuation dominant subspace and the hidden danger sensitive subspace are constructed respectively. The basis vectors of the hazard-sensitive subspace are orthogonalized to obtain the orthogonalized hazard-sensitive basis vectors. The vectors of the load fluctuation dominant subspace are concatenated with the orthogonalized hidden danger sensitive basis vectors to generate an adaptive dimensionality reduction mapping matrix.
5. The method according to claim 4, characterized in that, The process involves constructing a load fluctuation-dominant subspace and a potential hazard-sensitive subspace based on the normal sample set and the pre-stored early fatigue prior sample set of the gas valve spring, respectively, including: Obtain the global background features of the normal sample set and the local sensitive features of the pre-stored early fatigue prior sample set of the gas valve spring; The covariance matrix is calculated based on the global background features of the normal sample set, and eigenvalue decomposition is performed. The eigenvector corresponding to the largest eigenvalue is extracted, and the load fluctuation dominant subspace is constructed. The local covariance matrix is calculated based on the local sensitivity features of the pre-stored early fatigue prior sample set of the gas valve spring, and the main feature vector sensitive to spring degradation is extracted to construct the hidden danger sensitive subspace.
6. The method according to claim 1, characterized in that, The step of performing matrix multiplication on the high-dimensional initial feature vector and the preset adaptive dimensionality reduction mapping matrix to obtain the low-dimensional decoupled feature vector includes: The high-dimensional initial feature vector is multiplied by the load-dominant basis vector in the preset adaptive dimensionality reduction mapping matrix to obtain the first low-dimensional feature vector, wherein the first low-dimensional feature vector represents the current macroscopic load migration status of the device. The high-dimensional initial feature vector is multiplied by the hidden danger sensitive basis vector in the preset adaptive dimension reduction mapping matrix to obtain the second low-dimensional feature vector, wherein the second low-dimensional feature vector characterizes the degree of microscopic degradation of the local valve. The first low-dimensional feature vector and the second low-dimensional feature vector are oriented and projected to reconstruct the low-dimensional decoupled feature space, and the low-dimensional decoupled feature vector is output.
7. The method according to claim 1, characterized in that, The step of inputting the low-dimensional decoupled feature vector into the local normal baseline model to obtain the hidden danger index includes: A hypersphere is defined for the local normal baseline model, wherein the center and radius of the hypersphere characterize the normal data distribution under a specific load; The low-dimensional decoupling feature vector is input into the local normal benchmark model, and the Euclidean distance of the low-dimensional decoupling feature vector from the center of the hypersphere in the low-dimensional decoupling feature space is calculated. The hazard index is obtained based on the Euclidean distance, wherein the hazard index is positively correlated with the Euclidean distance.
8. The method according to claim 1, characterized in that, The calculation of the distance between the global background feature subset and each historical anchor point in the offline pre-constructed load state baseline manifold, and the generation of load background labels, includes: The historical fault-free data collected when the device undergoes various load adjustments under normal conditions is obtained. A subset of historical background features of the historical fault-free data is extracted and mapped to a low-dimensional embedding space through a manifold learning algorithm, so as to construct a load state reference manifold containing each historical anchor point and its corresponding embedding coordinates in the low-dimensional embedding space in an offline state. The global background feature subset is mapped to the low-dimensional embedding space. The embedding space distance between the mapped feature points and each historical anchor point in the load state baseline manifold is calculated. The K-nearest neighbor search algorithm is used to find the nearest historical normal load samples and obtain the embedding space distance value of the nearest neighbor samples. Obtain the embedding coordinates of each of the nearest neighbor samples in the low-dimensional embedding space, and perform a weighted summation of the embedding coordinates based on the reciprocal of the distance value in the embedding space as a weight to obtain the current embedding coordinate position of the global background feature subset in the low-dimensional embedding space; Based on the mapping relationship between each embedded coordinate position and the load level value pre-calibrated in the load state reference manifold, the current embedded coordinate position is converted into a load background label.
9. The method according to claim 1, characterized in that, The step of comparing the hazard index with the judgment threshold of the local normal benchmark model to obtain the discrimination result includes: When the hidden danger index is less than or equal to the judgment threshold, the judgment result is that the current state is normal; When the hidden danger index is greater than the judgment threshold, the offset vector is determined based on the difference between the low-dimensional decoupling feature vector and the standard center feature vector corresponding to the local normal benchmark model. The offset vector is orthogonally decomposed into a preset sensitive projection direction to obtain the offset component along the sensitive projection direction; When the offset component is less than or equal to the preset lower limit of sensitivity tolerance, the judgment result is normal operating condition drift; When the offset component is greater than the lower limit of the sensitivity tolerance, the judgment result is that the valve spring has undergone early fatigue degradation.
10. An intelligent industrial hazard identification system based on pattern recognition, characterized in that, include: The acquisition module is used to acquire multi-dimensional operating status data during the operation of the equipment, and to perform spatiotemporal synchronous acquisition and angular domain analysis on the multi-dimensional operating status data to determine local sensitive data windows and global background data windows. The multi-dimensional operating status data includes at least cylinder dynamic pressure signal, valve cover vibration acceleration signal and crankshaft key phase pulse signal. The extraction module is used to extract global background features from the global background data window and local sensitive features from the local sensitive data window. The extracted global background features and local sensitive features are concatenated and standardized to generate a high-dimensional initial feature vector. The calculation module is used to extract a global background feature subset from the high-dimensional initial feature vector, calculate the distance between the global background feature subset and each historical anchor point in the offline pre-constructed load state baseline manifold, and generate load background labels. The matrix multiplication module is used to perform matrix multiplication on the high-dimensional initial feature vector and the preset adaptive dimension reduction mapping matrix to obtain a low-dimensional decoupled feature vector, wherein the preset adaptive dimension reduction mapping matrix is obtained according to the load background label; The input module is used to call the corresponding local normal benchmark model from the pre-stored pattern matching model library according to the load background label, and input the low-dimensional decoupled feature vector into the local normal benchmark model to obtain the hidden danger index; The comparison module is used to compare the hidden danger index with the judgment threshold of the local normal benchmark model to obtain the judgment result.
