Real-time classification method and system based on multi-stage data feature slicing
Patent Information
- Application Number
- CN202611356319.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-10-30
AI Technical Summary
与当前故障预警无关的历史阶段特征，若仍参与模型推理，容易引入噪声特征并增加边缘计算设备的处理负载，从而导致状态识别延迟升高、误报率或漏报率增加
[0035]与现有技术相比，本发明通过采用单阶段独立处理机制，使每个阶段基于自身数据完成 特征提取、切片、建模和预测全流程，避免了多阶段数据相互依赖所导致的处理灵活性不足的问题，提升了各阶段特征利用的针对性。
Smart Images
Figure CN123012511A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of condition monitoring and fault early warning technology, and more specifically, to a real-time classification method and system for multi-stage data feature slices in the operation process of industrial equipment. Background Technology
[0002] In industrial production sites, industrial equipment such as motors, pumps, compressors, and CNC machine tools continuously generate operating data such as voltage, current, temperature, vibration, speed, pressure, or power during different operating stages, including startup, stable operation, high-load operation, and before shutdown. The equipment conditions differ at each stage, and their abnormal manifestations also vary. For example, abnormal current surges may occur during startup, abnormal temperature fluctuations may occur during stable operation, abnormal vibration peaks or excessively rapid temperature rise may occur during high-load operation, and abnormal parameter decay may occur before shutdown.
[0003] Existing industrial equipment condition identification methods typically input operational data from multiple operating stages into the classification model all at once, ignoring the differences in the contribution of different operating stages to the current fault state determination. If features from historical stages unrelated to the current fault warning are still included in model inference, they can easily introduce noise and increase the processing load on edge computing devices, leading to increased condition identification latency and higher false alarm or false negative rates.
[0004] While some industrial equipment monitoring solutions introduce multi-stage feature fusion mechanisms, they often use fixed starting stages for fusion, lacking dynamic control over the effective historical stage starting point. This results in each current stage requiring modeling and inference of data features from all historical stages. As the number of equipment operation stages or sensor channels increases, the feature dimensions accumulate continuously, easily leading to increased computational overhead at the edge and making it difficult to meet the response time requirements for real-time fault warnings in industrial settings.
[0005] Therefore, it is necessary to propose a real-time classification method for multi-stage data feature slices for industrial equipment condition monitoring and fault early warning. This method retains effective stage features related to the current equipment condition while eliminating outdated or irrelevant stage features, thereby reducing feature dimensions and computational complexity and improving the real-time performance and reliability of industrial equipment anomaly identification. Summary of the Invention
[0006] The purpose of this invention is to overcome the defects and shortcomings of the prior art and provide a real-time classification method and system based on multi-stage data feature slicing. Through stage data feature extraction, stage feature slicing, and stage-independent classification processing, the real-time identification of the operating status of industrial equipment and fault early warning can be achieved.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] A real-time classification method based on multi-stage data feature slicing includes the following steps:
[0009] Acquire operational data during the operation of industrial equipment and divide the operational data into multiple stages of operational data sequences that reflect changes in the state of industrial equipment;
[0010] Feature extraction is performed on the operational data sequence of each stage to obtain the equipment status feature set of each stage;
[0011] The feature set of equipment status at each stage is sliced to obtain the feature subset of each stage.
[0012] Each feature subset of each stage is input into the pre-trained device state classification model for classification prediction, and the device state classification results of each stage are obtained.
[0013] Output the equipment status classification results for each stage, and perform industrial equipment status monitoring or fault early warning based on the equipment status classification results for each stage.
[0014] Furthermore, the operating data includes at least one of voltage, current, temperature, vibration frequency, rotational speed, pressure, and power. According to the equipment operating conditions, time windows, or process flow nodes, the operating data is divided into multiple stages of operating data sequences that reflect changes in the state of industrial equipment.
[0015] Furthermore, the equipment status feature set includes at least one of statistical features, time-series variation features, and abnormal fluctuation features. Statistical features include mean, variance, or peak value. Time-series variation features include rate of change or trend slope. Abnormal fluctuation features include current fluctuation amplitude, vibration peak value, pressure fluctuation range, or power mutation amplitude.
[0016] Furthermore, the slicing process includes at least one of outdated feature vector removal and fixed-dimensional filtering.
[0017] Furthermore, the outdated feature vector removal process includes: determining the slice start index j, and removing features in the device status feature set that are located before the slice start index j and whose contribution to the current stage status recognition is lower than a preset threshold;
[0018] The slice starting index j is determined according to at least one of the following methods: based on the contribution of each stage feature to the equipment status classification result, based on the time interval between the feature generation time and the current discrimination time, based on the stability or abnormal correlation of the operating data, or based on the industrial equipment type or operating condition preset.
[0019] Furthermore, the fixed-dimensional filtering process includes: setting a fixed dimension v, which is dynamically configured based on the model computational complexity threshold, hardware computing power limitations, or classification accuracy requirements, and satisfies 1≤v≤ the total dimension of the device status features at the current stage, so that the dimension of the feature subset after slicing is no greater than v.
[0020] Furthermore, the slicing strategy is selected based on the current stage number k, the number of features w in the current stage device state feature set, the preset stage number threshold K0, and the preset dimension threshold dim0, specifically including:
[0021] Obtain the stage number k of the current pending stage and the number w of features in the current stage's device status feature set;
[0022] Set a threshold for the number of stages, K0, and a threshold for the number of dimensions, dim0, where K0 is the threshold for the number of stages to determine whether outdated features need to be removed, and dim0 is the threshold for the feature dimension to determine whether a fixed dimension is needed.
[0023] If k≤K0 and w≤dim0, then it is determined that the strategy of not using outdated feature vector removal and not using fixed dimension screening is not used.
[0024] If k≤K0 and w>dim0, then it is determined that the strategy of using only fixed dimensions for filtering is used;
[0025] If k>K0 and w≤dim0, then it is determined that only the strategy of removing outdated feature vectors is used;
[0026] If k>K0 and w>dim0, it is determined that the strategy of using both outdated feature vector removal and fixed-dimensional filtering is used simultaneously, and the fixed-dimensional filtering is performed after the outdated feature vector removal process.
[0027] Furthermore, the setting methods for the stage number threshold K0 and the dimension threshold dim0 include at least one of the following: adaptive adjustment based on the performance feedback results of historical classification data, preset fixed values based on the operating characteristics of industrial equipment, or dynamic configuration based on the application scenario's requirements for classification real-time performance and accuracy.
[0028] Furthermore, the equipment status classification model is at least one of support vector machine, random forest, K-nearest neighbor, or partial least squares. The equipment status classification model is trained based on historical labeled samples of the corresponding stage, and the training process uses cross-validation or grid search to optimize the model parameters.
[0029] A real-time classification system based on multi-stage data feature slices, applying any one of the above-described real-time classification methods based on multi-stage data feature slices, includes:
[0030] The data acquisition module is used to acquire operational data during the operation of industrial equipment and divide the operational data into multiple stages of operational data sequences that reflect changes in the state of the industrial equipment.
[0031] The feature extraction module is used to extract features from the operational data sequence of each stage to obtain the device status feature set of each stage.
[0032] The feature slicing module is used to perform feature slicing on the device state feature set of each stage to obtain the feature subset of each stage.
[0033] The stage classification module is used to input the feature subset of each stage into the pre-trained device state classification model, perform classification prediction, and obtain the device state classification results for each stage.
[0034] The results output module is used to output the equipment status classification results for each stage, and to perform industrial equipment status monitoring or fault early warning based on the equipment status classification results for each stage.
[0035] Compared with existing technologies, this invention adopts a single-stage independent processing mechanism, enabling each stage to complete the entire process of feature extraction, slicing, modeling and prediction based on its own data. This avoids the problem of insufficient processing flexibility caused by the interdependence of data in multiple stages and improves the targeting of feature utilization in each stage.
[0036] This invention performs slicing and classification operations independently at each stage, and can output the current stage results in real time without waiting for data from subsequent stages. This reduces the response latency from the traditional full-stage processing time to the single-stage processing time, significantly improving the real-time performance of stage status feedback.
[0037] This invention improves the fitting accuracy of single-stage features by using stage feature vectorization and independent modeling of each stage, accurately depicting the state differences of each stage using an independent feature space, and combining independent model parameter configurations for each stage, thereby enhancing the accuracy of stage classification results.
[0038] Each module of the system of this invention is set up in a one-to-one correspondence with the data of the operation stage. The feature extraction, slicing strategy and classification model can be adjusted according to different types of industrial equipment, sensor configurations and operating conditions. It is widely applicable to online status monitoring, anomaly identification and fault early warning scenarios of industrial equipment. Attached Figure Description
[0039] Figure 1 This is a flowchart of a real-time classification method based on multi-stage data feature slicing.
[0040] Figure 2 This is a schematic diagram of a real-time classification system based on multi-stage data feature slicing. Detailed Implementation
[0041] The real-time classification method and system based on multi-stage data feature slicing of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0042] Please see Figure 1 This invention discloses a real-time classification method based on multi-stage data feature slices, comprising the following steps:
[0043] S1: Acquire operational data during the operation of industrial equipment and divide the operational data into multiple operational data sequences that reflect changes in the state of industrial equipment;
[0044] S2: Extract features from the operational data sequence of each stage to obtain the equipment status feature set for each stage;
[0045] S3: Perform feature slicing on the equipment state feature set of each stage to obtain the feature subset of each stage;
[0046] S4: Input the feature subset of each stage into the pre-trained device state classification model, perform classification prediction, and obtain the device state classification results for each stage.
[0047] S5: Output the equipment status classification results for each stage, and perform industrial equipment status monitoring or fault early warning based on the equipment status classification results for each stage.
[0048] In step S1, the operating data includes voltage, current, temperature, vibration frequency, rotational speed, pressure, and power. According to the equipment operating conditions, time windows, or process flow nodes, the operating data is divided into multiple stages of operating data sequences that reflect changes in the state of industrial equipment.
[0049] Specifically, the operational data of industrial equipment during operation is acquired, and the operational data set is defined as follows:
[0050] ;
[0051] Where m represents the total number of industrial devices, and the operating data sequence of a single industrial device is as follows:
[0052] ;
[0053] Where n is the dimension of the running data, and the dimension of the running data in each stage is n;
[0054] The running data is divided into multiple stages according to time sequence or processing flow, denoted as stage 1, stage 2, ..., stage k, where k is a positive integer. The running data sequence of multiple stages is represented as follows:
[0055] ;
[0056] in, This represents the nth dimension of the operating data of industrial equipment i in the kth stage.
[0057] Each stage corresponds to a specific time window or processing interval, which is used to reflect the state change characteristics of industrial equipment within that stage.
[0058] In this embodiment, the application scenario is based on the classification of industrial equipment operating status. Operating data from 1000 similar production devices are acquired, and the set of operating data is denoted as […]. The operating data sequence of a single device is 12 of these are operational characteristic dimensions, including voltage, current, temperature, vibration frequency, etc.
[0059] Based on the time sequence of equipment operation, the operational data is divided into four stages, and the operational data sequence of the four stages is represented as follows:
[0060] .
[0061] In step S2, the equipment status feature set includes statistical features, time-series change features, and abnormal fluctuation features. Statistical features include mean, variance, or peak value. Time-series change features include rate of change or trend slope. Abnormal fluctuation features include current fluctuation amplitude, vibration peak value, pressure fluctuation range, or power mutation amplitude.
[0062] Specifically, feature extraction is performed on the operational data sequence for each stage to obtain the equipment state feature set for each stage, forming the corresponding stage feature vector:
[0063] , ;
[0064] The feature extraction process follows the characteristics of data distribution to ensure that the features can represent the stage state. The types of features extracted include, but are not limited to, statistical features, time-series change features, or abnormal fluctuation features.
[0065] The mathematical expression for a single-stage feature vector is:
[0066] ;
[0067] in, Let w represent the w-th feature extracted in the k-th stage, where w = c(k) is the feature dimension of the k-th stage, and c(k) is the feature dimension function corresponding to stage k, which can be dynamically adjusted according to the data characteristics.
[0068] This staged feature modeling approach uses vector quantization to characterize the distribution differences of data at different stages in the feature space, providing structured input for subsequent independent classification.
[0069] In this embodiment, feature extraction is performed on the operational data of each stage. The extracted feature types include statistical features and time-series variation features. Statistical features include mean, variance, or peak value, while time-series variation features include rate of change or trend slope.
[0070] Phase 1: Startup phase, extracting 7 features, feature vector is , This corresponds to key characteristics such as startup voltage stability and startup time;
[0071] Phase 2: Stable operation phase, extracting 9 features, feature vector is , This corresponds to characteristics such as temperature fluctuations and current stability during stable operation;
[0072] Phase 3: High-load phase, extracting 8 features, feature vector is , This corresponds to characteristics such as vibration peak value and power consumption under high load;
[0073] Phase 4: Pre-shutdown phase, extracting 6 features, feature vector is , This corresponds to the characteristics of parameter decay rate and voltage drop magnitude before shutdown.
[0074] In step S3, the slicing process includes outdated feature vector removal and fixed-dimensional filtering.
[0075] The outdated feature vector removal process includes: determining the slice start index j, and removing features in the device status feature set that are located before slice start index j and whose contribution to the current stage status identification is lower than a preset threshold. Outdated features are those whose contribution to the current device status classification is lower than a preset threshold, exceed a preset effective time window, or are not directly related to the current operating condition.
[0076] The slice starting index j is determined based on the following methods: the contribution of each stage's features to the equipment status classification result, the time interval between the feature generation time and the current judgment time, the stability or abnormal correlation of the operating data, or a preset industrial equipment type or operating condition. When the number of stages is small or the feature dimension does not exceed the preset threshold, outdated feature vector removal processing may not be performed. When the number of stages is large or the feature dimension exceeds the preset threshold, outdated feature vector removal processing is performed.
[0077] The fixed-dimensional filtering process includes: setting a fixed dimension v, which is dynamically configured based on the model computational complexity threshold, hardware computing power limitations, or classification accuracy requirements, and satisfies 1≤v≤ the total dimension of the device status features at the current stage, so that the dimension of the feature subset after slicing is no greater than v.
[0078] The slicing strategy is selected based on the current stage number k, the number of features w in the current stage's device state feature set, the preset stage number threshold K0, and the preset dimension threshold dim0, specifically including:
[0079] Obtain the stage number k of the current pending stage and the number w of features in the current stage's device status feature set;
[0080] Set a threshold for the number of stages, K0, and a threshold for the number of dimensions, dim0, where K0 is the threshold for the number of stages to determine whether outdated features need to be removed, and dim0 is the threshold for the feature dimension to determine whether a fixed dimension is needed.
[0081] If k≤K0 and w≤dim0, then it is determined that the strategy of not using outdated feature vector removal and not using fixed dimension screening is not used.
[0082] If k≤K0 and w>dim0, then it is determined that the strategy of using only fixed dimensions for filtering is used;
[0083] If k>K0 and w≤dim0, then it is determined that only the strategy of removing outdated feature vectors is used;
[0084] If k>K0 and w>dim0, it is determined that the strategy of using both outdated feature vector removal and fixed-dimensional filtering is used simultaneously, and the fixed-dimensional filtering is performed after the outdated feature vector removal process.
[0085] The methods for setting the stage number threshold K0 and the dimension threshold dim0 include: adaptive adjustment based on the performance feedback results of historical classification data, preset fixed values based on the operating characteristics of industrial equipment, or dynamic configuration based on the application scenario's requirements for classification real-time performance and accuracy.
[0086] Specifically, the feature vector of each stage is extracted. The feature slicing operation is performed independently, and the slicing process is implemented using outdated feature vector removal or fixed-dimensional filtering. Let the starting index of the slice be j, and the number of feature vectors in the current stage be t. Then the feature subset after slicing in the k-th stage is:
[0087] ;
[0088] Where j is the starting index of the slice, satisfying Each stage requires slicing. By using outdated feature vector removal or fixed-dimensional slicing methods, the number and dimension of feature vectors in the feature subset of each stage can be controlled, thereby controlling the size of the feature subset at each stage, which helps in implementing classification tasks.
[0089] In this embodiment, the feature vector of each stage is sliced independently. The slicing process adopts a fixed-dimensional filtering method, and according to the principle of "prioritizing core features and discarding features that do not meet the requirements", a subset of 5-dimensional effective features is uniformly extracted, as follows:
[0090] In the first stage, with feature vector dimension w=7: the top 5 core effective features are selected through feature importance evaluation, while the last 2 secondary features are discarded, resulting in a feature subset. ;
[0091] In the second stage, with feature vector dimension w=9, redundant or noisy features in the four dimensions that do not meet the requirements for representing the stage state are removed, and the middle five key and effective features are retained, forming a feature subset. ;
[0092] In the third stage, with feature vector dimension w=8, the last three dimensions of features that do not meet the requirements for high-load state determination are discarded, and the first five core effective features are extracted, forming a feature subset. ;
[0093] In stage 4, with feature vector dimension w=7: the last dimension, which contributes nothing to the prediction of the shutdown state, is discarded, and the first 5 effective features are retained, forming a feature subset. .
[0094] In step S4, the equipment status classification model is a support vector machine, random forest, K-nearest neighbors or partial least squares. The equipment status classification model is trained based on historical labeled samples of the corresponding stage. The training process uses cross-validation or grid search to optimize the model parameters.
[0095] Specifically, the feature subset St after each stage slice is independently input into the classification model to complete the classification prediction for that stage. The prediction process of the classification model can be represented as follows:
[0096] ;in For classification models, θ represents the model parameters. The classification result at stage t represents the classification model, which can be support vector machine, random forest, K-nearest neighbors, partial least squares, or a combination thereof. The model at each stage is the optimal model selected after a fixed number of grid search iterations. The performance of the model needs to be evaluated by comprehensively considering indicators such as accuracy, specificity, sensitivity, and Matthews correlation coefficient.
[0097] Each indicator has a different focus: accuracy focuses on the proportion of correct predictions overall, reflecting the model's comprehensive classification ability; specificity focuses on the correct identification rate of "normal samples (negative)" to avoid misclassifying normal samples as abnormal; sensitivity focuses on the correct identification rate of "abnormal samples (positive)" to reduce the risk of missing abnormal samples; and the Matthews correlation coefficient balances the effects of true positives, true negatives, false positives, and false negatives, making it a more reliable comprehensive performance measure under imbalanced samples.
[0098] The independent classification modeling and prediction operations described above are performed in stages 1 through k, namely:
[0099] Phase 1: Phase 2: ...; Phase k: .
[0100] In this embodiment, the effective historical stage range is determined based on the current stage. The effective historical stage and the equipment status characteristics of the current stage are fused, and the fused feature data is input into the equipment status classification model to output the equipment operating status and fault warning results of the current stage. The classification model used in each stage is a model optimized through grid search or cross-validation, as detailed below:
[0101] Phase 1: A random forest model was selected, with model parameters set to 100 decision trees and a maximum depth of 8. The prediction process was as follows: Output classification results 0 indicates normal startup, and 1 indicates abnormal startup;
[0102] Phase 2: A Support Vector Machine (SVM) model is selected, with the kernel function set to RBF and a penalty coefficient C=1.0. The prediction process is as follows: Output classification results 0 indicates stable operation, and 1 indicates fluctuating operation;
[0103] Phase 3: Partial least squares model is selected, and the prediction process is as follows: Output classification results 0 indicates normal high load, and 1 indicates high load overload;
[0104] Phase 4: Using the K-nearest neighbor model with a neighborhood number of k=5, the prediction process is as follows: Output classification results 0 indicates normal shutdown, and 1 indicates abnormal shutdown.
[0105] The feature slicing process here uses a fixed dimension, which effectively reduces the risk of "dimensionality explosion". If the processing stage is longer, the outdated feature vectors should be processed first before the fixed dimension is applied.
[0106] Step S5: Output the equipment status classification results for each stage, and perform industrial equipment status monitoring or fault early warning based on the equipment status classification results for each stage.
[0107] Specifically, the equipment status classification results for each stage are output separately. Let l = 1, 2, ..., k, and let l be the value of the classifier. This classifier is used for industrial equipment condition monitoring or fault early warning. For multi-classification scenarios, the output results of each stage can be represented as a probability distribution:
[0108] ;
[0109] Where C is the total number of categories, satisfying .
[0110] In this embodiment, the equipment status classification results for each stage are output, including the status determination, probability distribution, and confidence level for each stage:
[0111] Phase 1 output: Feedback on the device startup status;
[0112] Phase 2 output: Feedback on the stable operation status of the equipment;
[0113] Phase 3 output: Feedback on the status of the equipment during high-load periods;
[0114] Phase 4 output: It provides feedback on the status of the equipment before it shuts down.
[0115] Please see Figure 2 The present invention also discloses a real-time classification system based on multi-stage data feature slices, which applies the real-time classification method based on multi-stage data feature slices described in any of the above claims, including:
[0116] The data acquisition module is used to acquire operational data during the operation of industrial equipment and divide the operational data into multiple stages of operational data sequences that reflect changes in the state of the industrial equipment.
[0117] The feature extraction module is used to extract features from the operational data sequence of each stage to obtain the device status feature set of each stage.
[0118] The feature slicing module is used to perform feature slicing on the device state feature set of each stage to obtain the feature subset of each stage.
[0119] The stage classification module is used to input the feature subset of each stage into the pre-trained device state classification model, perform classification prediction, and obtain the device state classification results for each stage.
[0120] The results output module is used to output the equipment status classification results for each stage, and to perform industrial equipment status monitoring or fault early warning based on the equipment status classification results for each stage.
[0121] The output of the data acquisition module is connected to the input of the feature extraction module in a one-to-one correspondence; the output of the feature extraction module is connected to the input of the feature slicing module in a one-to-one correspondence; the output of the feature slicing module is connected to the input of the stage classification module in a one-to-one correspondence; and the output of the stage classification module is connected to the input of the result output module in a one-to-one correspondence. The modules interact with each other through a data flow interface to ensure that each stage independently completes the entire process of "running data acquisition → feature extraction → slicing → modeling → status recognition" and realizes industrial equipment status monitoring and fault early warning based on the independent classification results of each stage.
[0122] In summary, this invention employs a single-stage independent processing mechanism, enabling each stage to complete the entire process of feature extraction, slicing, modeling, and prediction based on its own data. This avoids the problem of insufficient processing flexibility caused by the interdependence of data in multiple stages and improves the targeted utilization of features at each stage.
[0123] This invention performs slicing and classification operations independently at each stage, and can output the current stage results in real time without waiting for data from subsequent stages. This reduces the response latency from the traditional full-stage processing time to the single-stage processing time, significantly improving the real-time performance of stage status feedback.
[0124] This invention improves the fitting accuracy of single-stage features by using stage feature vectorization and independent modeling of each stage, accurately depicting the state differences of each stage using an independent feature space, and combining independent model parameter configurations for each stage, thereby enhancing the accuracy of stage classification results.
[0125] Each module of the system of this invention is set up in a one-to-one correspondence with the data of the operation stage. The feature extraction, slicing strategy and classification model can be adjusted according to different types of industrial equipment, sensor configurations and operating conditions. It is widely applicable to online status monitoring, anomaly identification and fault early warning scenarios of industrial equipment.
[0126] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit disclosed in the present invention should fall within the patent scope covered by the present invention.
Claims
1. A real-time classification method based on multi-stage data feature slicing, characterized in that, Includes the following steps: Acquire operational data during the operation of industrial equipment and divide the operational data into multiple stages of operational data sequences that reflect changes in the state of industrial equipment; Feature extraction is performed on the operational data sequence of each stage to obtain the equipment status feature set of each stage; The feature set of equipment status at each stage is sliced to obtain the feature subset of each stage. Each feature subset of each stage is input into the pre-trained device state classification model for classification prediction, and the device state classification results of each stage are obtained. Output the equipment status classification results for each stage, and perform industrial equipment status monitoring or fault early warning based on the equipment status classification results for each stage.
2. The real-time classification method based on multi-stage data feature slicing according to claim 1, characterized in that, Operating data includes at least one of voltage, current, temperature, vibration frequency, rotational speed, pressure, and power. According to the equipment operating conditions, time windows, or process flow nodes, the operating data is divided into multiple stages of operating data sequences that reflect changes in the state of industrial equipment.
3. The real-time classification method based on multi-stage data feature slicing according to claim 1, characterized in that, The equipment status feature set includes at least one of statistical features, time-series variation features, and abnormal fluctuation features. Statistical features include mean, variance, or peak value. Time-series variation features include rate of change or trend slope. Abnormal fluctuation features include current fluctuation amplitude, vibration peak value, pressure fluctuation range, or power sudden change amplitude.
4. The real-time classification method based on multi-stage data feature slicing according to claim 1, characterized in that, Slicing processing includes at least one of outdated feature vector removal and fixed-dimensional filtering.
5. The real-time classification method based on multi-stage data feature slicing according to claim 4, characterized in that, The outdated feature vector removal process includes: determining the slice start index j, and removing features in the device status feature set that are located before the slice start index j and whose contribution to the current stage status recognition is lower than a preset threshold. The slice starting index j is determined according to at least one of the following methods: based on the contribution of each stage feature to the equipment status classification result, based on the time interval between the feature generation time and the current discrimination time, based on the stability or abnormal correlation of the operating data, or based on the industrial equipment type or operating condition preset.
6. The real-time classification method based on multi-stage data feature slicing according to claim 4, characterized in that, The fixed-dimensional filtering process includes: setting a fixed dimension v, which is dynamically configured based on the model computational complexity threshold, hardware computing power limitations, or classification accuracy requirements, and satisfies 1≤v≤ the total dimension of the device status features at the current stage, so that the dimension of the feature subset after slicing is no greater than v.
7. The real-time classification method based on multi-stage data feature slicing according to claim 1, characterized in that, The slicing strategy is selected based on the current stage number k, the number of features w in the current stage's device state feature set, the preset stage number threshold K0, and the preset dimension threshold dim0, specifically including: Obtain the stage number k of the current pending stage and the number w of features in the current stage's device status feature set; Set a threshold for the number of stages, K0, and a threshold for the number of dimensions, dim0, where K0 is the threshold for the number of stages to determine whether outdated features need to be removed, and dim0 is the threshold for the feature dimension to determine whether a fixed dimension is needed. If k≤K0 and w≤dim0, then it is determined that the strategy of not using outdated feature vector removal and not using fixed dimension screening is not used. If k≤K0 and w>dim0, then it is determined that the strategy of using only fixed dimensions for filtering is used; If k>K0 and w≤dim0, then it is determined that only the strategy of removing outdated feature vectors is used; If k>K0 and w>dim0, it is determined that the strategy of using both outdated feature vector removal and fixed-dimensional filtering is used simultaneously, and the fixed-dimensional filtering is performed after the outdated feature vector removal process.
8. The real-time classification method based on multi-stage data feature slicing according to claim 7, characterized in that, The setting methods for the stage number threshold K0 and the dimension threshold dim0 include at least one of the following: adaptive adjustment based on the performance feedback results of historical classification data, preset fixed values based on the operating characteristics of industrial equipment, or dynamic configuration based on the application scenario's requirements for classification real-time performance and accuracy.
9. The real-time classification method based on multi-stage data feature slicing according to claim 1, characterized in that, The equipment status classification model is at least one of support vector machine, random forest, K-nearest neighbor or partial least squares. The equipment status classification model is trained based on historical labeled samples of the corresponding stage. The training process uses cross-validation or grid search to optimize the model parameters.
10. A real-time classification system based on multi-stage data feature slicing, characterized in that, The real-time classification method based on multi-stage data feature slices according to any one of claims 1 to 9 includes: The data acquisition module is used to acquire operational data during the operation of industrial equipment and divide the operational data into multiple stages of operational data sequences that reflect changes in the state of the industrial equipment. The feature extraction module is used to extract features from the operational data sequence of each stage to obtain the device status feature set of each stage. The feature slicing module is used to perform feature slicing on the device state feature set of each stage to obtain the feature subset of each stage. The stage classification module is used to input the feature subset of each stage into the pre-trained device state classification model, perform classification prediction, and obtain the device state classification results for each stage. The results output module is used to output the equipment status classification results for each stage, and to perform industrial equipment status monitoring or fault early warning based on the equipment status classification results for each stage.