Equipment fault detection method based on dynamic pyramid

Abstract

The invention provides an equipment fault detection method based on a dynamic pyramid, and the method comprises the steps: collecting and preprocessing multi-source heterogeneous data, and generating a standardized time sequence feature set; a dynamic pyramid model is constructed, and features are aggregated through multi-layer time granularity; adaptively adjusting a pyramid structure based on a characteristic volatility index, and optimizing time scale characterization; generating a global equipment state feature vector through an adaptive feature weighted fusion mechanism; a fault diagnosis optimization mechanism based on a dobby machine is adopted, a Thompson sampling algorithm is utilized to dynamically select a fault mode, and a confidence index is introduced to evaluate the diagnosis reliability. According to the method, through the dynamic pyramid model, adaptive feature fusion and an online optimization strategy, the refinement of feature representation, the dynamic adaptability of the model and the long-term accuracy of diagnosis are remarkably improved, and the method is suitable for equipment fault detection under complex working conditions and has the technical effects of high efficiency, robustness and reliability.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent device management, and in particular relates to a device fault detection method based on a dynamic pyramid. Background Art

[0002] With the rapid development of emerging technologies such as the Industrial Internet and intelligent manufacturing, the scale and complexity of various complex equipment systems are constantly increasing, generating massive amounts of multi-source, heterogeneous data during their operation. How to fully tap the value of this data and achieve accurate diagnosis and prediction of equipment failures has become a key issue that needs to be addressed.

[0003] Traditional equipment fault detection methods are primarily based on single data sources (such as vibration and temperature) and employ expert experience to build shallow models. These methods are unable to cope with dynamic changes under complex operating conditions. Existing equipment fault detection solutions have at least the following drawbacks:

[0004] (1) Simple feature splicing or decision weighting is usually used for multi-source heterogeneous data, ignoring the correlation and complementarity between different data sources, resulting in shallow feature fusion and insufficient diagnostic accuracy;

[0005] (2) Existing methods mostly use fixed shallow models or static feature extraction frameworks, which lack adaptability to the dynamic evolution of equipment status and cannot effectively cope with time scale changes under complex working conditions, resulting in poor diagnostic robustness;

[0006] (3) Existing fault diagnosis methods mostly rely on offline training models and lack online learning capabilities. They are unable to dynamically optimize diagnostic strategies to adapt to real-time changes in equipment status, resulting in a decline in long-term diagnostic performance.

[0007] Therefore, we need to develop an equipment fault detection method based on a dynamic pyramid, which can improve the accuracy of fault diagnosis, enhance dynamic adaptability, achieve long-term optimization, and ensure the efficiency and reliability of fault detection under complex working conditions. Summary of the Invention

[0008] The purpose of the present invention is to provide a device fault detection method based on a dynamic pyramid to solve the problems of shallow feature fusion, rigid diagnosis model, lack of online learning, etc. in the existing equipment fault detection scheme mentioned in the above background technology when facing multi-source heterogeneous big data.

[0009] To achieve the above objectives, the present invention provides a device fault detection method based on a dynamic pyramid, the method being as follows:

[0010] Multi-source heterogeneous data of the equipment is collected and preprocessed to obtain a multi-source time series feature set, and the multi-source time series feature set is aggregated at L different time granularities to obtain L layers of aggregated features and construct an initial dynamic pyramid model.

[0011] Based on the characteristic volatility index, adaptively adjusting the number of layers and time granularity of the dynamic pyramid model to generate an adaptive dynamic pyramid;

[0012] The adaptive adjustment requires first adjusting each layer of the dynamic pyramid model , calculate the arithmetic mean of the characteristic volatility index of all its aggregated features as the average volatility intensity, and the calculation formula is as follows: ,in For each layer The number of aggregated features in , is the timestamp, is the characteristic volatility indicator, is the average volatility intensity;

[0013] Based on the aggregated features of the adaptive dynamic pyramid, an adaptive feature weighted fusion mechanism is adopted to generate a global device state feature vector;

[0014] Based on the multi-armed bandit machine, the fault diagnosis optimization of the device state feature vector is performed through the Thompson sampling algorithm, and the failure mode and confidence index of the device are output.

[0015] Based on the above solution, the initial dynamic pyramid model construction includes:

[0016] Based on the time series feature set, a The dynamic pyramid model of the layers The layer time granularity is , is the sampling interval, , the time granularity increases from the bottom layer to the top layer;

[0017] Perform aggregation operations on each time series feature at each layer to generate the first Tier Aggregate features , aggregate each feature exist The results of aggregation at each granularity level constitute the initial dynamic pyramid model : .

[0018] Based on the above scheme, for the pyramid Layer Aggregate features , its volatility index Defined as the aggregate feature in the historical time window coefficient of variation within ;

[0019] The adaptive adjustment is based on the minimum preset threshold after calculating the average volatility intensity. and the highest preset threshold , determine the average volatility index of each layer Do they all fall within the interval? If not, adjust the number of layers and time granularity through splitting or merging operations until the average volatility indicators of all layers fall within the interval , and obtain the adaptive dynamic pyramid.

[0020] Based on the above solution, the splitting operation includes:

[0021] When Layer average volatility intensity , in Layer and A split aggregation layer is inserted between the layers, and its time granularity is Layer and Arithmetic mean of layer time granularity: , and based on the time granularity, the aggregation characteristics of the split aggregation layer are generated as follows: ;

[0022] The merging operation includes:

[0023] Jordi Layer average volatility intensity , and the average volatility of the previous layer Close, that is: ,in is a small positive threshold, then Layer and Layer merging, generating time granularity is The combined aggregation layer of , based on time granularity Regenerate aggregate features of merged aggregate layers .

[0024] Based on the above solution, the adaptive feature weighted fusion includes:

[0025] Based on the Layer Aggregate features Characteristic volatility indicator , define its fusion weight as: , in, , is a hyperparameter that controls the degree of concentration of fusion weight distribution;

[0026] Based on the fusion weight, layer Aggregate features are weighted fused to generate the Layer fusion features: , and generate the global device state feature vector through global weighted fusion ,in For the The global weight of the layer fusion features, Indicates the total number of layers of the adaptive dynamic pyramid.

[0027] Based on the above scheme, the global weighted weight is set to Layer Average Volatility Index The normalized value of , that is: ,in is the layer number index variable, and its value range is from the 1st layer to the layers, so that the layers with higher volatility occupy a greater weight in the global fusion, and enhance the ability to represent dynamic changes.

[0028] Based on the above solution, the fault diagnosis optimization includes: offline construction including Feature dictionary of known equipment failure modes , where each The characteristic vector corresponding to a certain fault mode and the characteristic vector of the equipment state Same dimension;

[0029] Online calculation of device state feature vectors Characteristics of each failure mode Similarity: ;

[0030] Define expected return as: , the equipment fault diagnosis problem is formalized as a multi-armed bandit problem.

[0031] Based on the above scheme, the Thompson sampling algorithm includes:

[0032] For each failure mode The maintenance parameters are and Beta distribution , initial settings , ;

[0033] At time step , estimated returns from various Beta distributions : , select the failure mode with the largest estimated benefit value , , and the failure mode The corresponding eigenvector , as the current diagnosis result;

[0034] Update parameters according to the actual device status: If the current diagnosis result is correct, increase ,make: , otherwise increase ,make: , iteratively optimize the diagnostic strategy.

[0035] Based on the above scheme, the confidence index is defined as: , used to evaluate the reliability of the current diagnostic results, A value close to 1 indicates that the current diagnosis result closely matches the device status.

[0036] Based on the above solution, the similarity measurement method includes Euclidean distance, cosine similarity or Pearson correlation coefficient, and the multi-source heterogeneous data includes sensor data, equipment operation logs and environmental parameters.

[0037] The present invention has the following advantages and effects compared to the prior art:

[0038] (1) Based on the volatility index of aggregated features, a weight is dynamically assigned to each feature in the form of a normalized exponential function, and global weighted fusion is performed between layers to form a global equipment status feature vector. This deeply mines the correlation and dynamic change trend of multi-source heterogeneous data, improves the refinement and enrichment of feature representation, and thus greatly improves the accuracy of fault diagnosis. It is suitable for equipment status monitoring under complex working conditions.

[0039] (2) A dynamic pyramid is constructed through multiple layers of time granularity, and the number of layers and granularity are dynamically adjusted based on the characteristic volatility index. The model structure is optimized through splitting or merging operations. The adaptive adjustment mechanism of the dynamic pyramid model can autonomously optimize the time granularity according to the volatility of the equipment status, accurately capture multi-scale dynamic characteristics, and significantly enhance the adaptability to complex working conditions and diagnostic robustness.

[0040] (3) A fault diagnosis optimization mechanism based on a multi-armed bandit is introduced, and the Thompson sampling algorithm is used to dynamically select the fault mode. The algorithm maintains the Beta distribution of each fault mode, samples the estimated benefits and updates the parameters at each time step to achieve a balance between exploration and utilization, and introduces a confidence indicator to evaluate the reliability of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly describes the drawings required for the specific embodiments or the description of the prior art. Similar elements or parts are generally identified by similar reference numerals throughout the drawings. Elements or parts in the drawings are not necessarily drawn to scale.

[0042] Figure 1 This is a flow chart of a device fault detection method based on a dynamic pyramid provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to more clearly illustrate the purpose, technical solutions and advantages of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The example implementation methods can be implemented in various forms and should not be understood as being limited to the examples described herein. On the contrary, these implementation methods are provided to make the present invention more comprehensive and complete, and to fully convey the concepts of the example implementation methods to those skilled in the art.

[0044] In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present invention. However, it will be appreciated by those skilled in the art that the technical solutions of the present invention can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring various aspects of the present invention.

[0045] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0046] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, while others may be combined or partially combined. Therefore, the actual execution order may vary depending on the actual situation.

[0047] The present invention will be described in detail below with reference to specific embodiments:

[0048] As attached Figure 1 As shown, embodiment 1 of the present invention provides a device fault detection method based on a dynamic pyramid, and the specific steps of the method are as follows:

[0049] Step 1: Multi-source heterogeneous data collection and preprocessing

[0050] First, collect the multi-source heterogeneous data generated during the operation of the equipment, including sensor data, equipment operation logs, and environmental parameters. The time series characteristics generated by the data source are ,in , is the total number of data sources, is the timestamp.

[0051] For time series features Perform preprocessing operations including cleaning, time alignment, and missing value interpolation to obtain a standardized time series feature set .in, After preprocessing, A time series feature.

[0052] Then, Perform maximum and minimum value normalization:

[0053]

[0054] in, is the normalized time series feature, recorded as the normalized time series feature, and its value range is [0, 1].

[0055] Then, yes Perform exponentially weighted average smoothing to remove high-frequency noise:

[0056]

[0057] in, is the smoothing coefficient, is the smoothed time series feature, recorded as smoothed time series feature.

[0058] After the above processing, the standardized multi-source time series feature set is obtained .

[0059] Step 2: Dynamic pyramid model construction

[0060] Based on the multi-source time series feature set , build a The dynamic pyramid model is a layered dynamic pyramid model. The time granularity of the layer is , and satisfies , where level number From the bottom to the top of the pyramid, the Corresponding to the lowest level (the finest time granularity), Corresponding to the top layer (coarsest time granularity). Initially, let ,in is the sampling interval of the multi-source time series feature set.

[0061] Specifically, for each multi-source timing feature , performing aggregation operations at L different time granularities , get L-layer aggregate features, where the first Tier The aggregate features are:

[0062]

[0063] The aggregation operation Including arithmetic mean, geometric mean, maximum value, and minimum value.

[0064] Each aggregate feature exist The results of aggregation at each granularity level constitute the initial dynamic pyramid model :

[0065]

[0066] in, is a The matrix describes the state of the device in Dynamic characteristics at different time scales.

[0067] Step 3: Adaptive structural adjustment of the dynamic pyramid model

[0068] Based on the initial dynamic pyramid model , combined with an adaptive structural adjustment mechanism, can better capture the dynamic characteristics of the device state at different time scales. The core of the structural adjustment mechanism is to use a characteristic volatility index , quantify the fluctuation degree of the aggregated features, and adaptively adjust the number of layers L and the time granularity of the dynamic pyramid model based on the fluctuation degree.

[0069] Specifically, for the pyramid Layer Aggregate features , its volatility index Defined as the aggregate feature in the historical time window The coefficient of variation within the historical time window , the length is , then the characteristic volatility index is:

[0070]

[0071] in is a small regularization term used to prevent zero division anomalies when the mean is small. and Represents the aggregation features In the historical time window The mean and standard deviation within , that is:

[0072] , ;

[0073] It should be noted that the volatility index Reflects the Tier Aggregate features in the history window The intensity of fluctuations within The larger the value, the more aggregated the features are. On the contrary, the fluctuation is relatively stable.

[0074] Furthermore, based on the characteristic volatility index, the structure of the dynamic pyramid model is adaptively adjusted.

[0075] First, for each layer of the dynamic pyramid model , calculate the volatility index of all its aggregate characteristics The arithmetic mean of , as the average volatility intensity, is calculated as follows:

[0076]

[0077] Then, based on the lowest preset threshold and the highest preset threshold , the average volatility index for each layer Make the following judgment:

[0078] like , it indicates that The aggregated features of the layer fluctuate violently as a whole and need to be further subdivided to capture finer-grained dynamic features. Layer and A split aggregation layer is inserted between the layers, and its time granularity is Take the first Layer and Arithmetic mean of layer size: , and based on granularity The multi-source time series features are re-aggregated to generate the aggregated features of the split aggregation layer: .

[0079] like , and the average volatility of the previous layer Close, that is:

[0080]

[0081] in is a small positive threshold. Layer and The layers are merged to form a coarser-grained aggregate layer with a time granularity of Take it as: , based on time granularity Regenerate aggregate features of merged aggregate layers .

[0082] The above splitting and merging operations are performed iteratively until the average volatility index of all layers is All fall within the interval Within, that is, a certain equilibrium state is reached, and the total number of layers of the dynamic pyramid model after adaptive adjustment is , its ℓth layer time granularity The aggregated features are , forming the adjusted dynamic pyramid model: , denoted as adaptive dynamic pyramid.

[0083] Through adaptive structural adjustments, the dynamic pyramid model can autonomously refine or coarsen the modeling granularity based on the volatility of aggregated features, thereby more accurately and comprehensively depicting the dynamic evolution of equipment status over time. This adaptive adjustment mechanism empowers the model to cope with complex operating conditions, enabling it to promptly capture key abnormal characteristics even in the face of step-like changes in equipment status, providing a reliable basis for subsequent fault diagnosis.

[0084] It should be noted that the threshold 、 as well as The value of needs to be set according to the specific application scenario and experience. Generally speaking, and This can be determined by statistical analysis of a large amount of historical data, such as taking all Value Percentiles (e.g. ) as ,Pick Percentile (such as 80%) as ;and It can be set to and In practical applications, these thresholds can be fine-tuned through experience to achieve the best diagnostic performance.

[0085] Step 4: Adaptive feature weighted fusion

[0086] It should be noted that after the adaptive structure adjustment in step 3, the adaptive dynamic pyramid It has been able to capture the dynamic characteristics of the device state at different time scales. However, the characteristics at different time scales may have different representation capabilities for the device state. In order to further highlight the features with strong representation capabilities and suppress the influence of redundant or noisy features, this paper introduces an adaptive feature weighted fusion mechanism based on this. The core of this mechanism is: based on the feature volatility index of each aggregated feature , constructing an adaptive weighted fusion framework so that features with greater volatility (i.e., those that carry more dynamic information about the device state) receive greater weight during fusion; conversely, features with less volatility receive relatively lower weights. This fused feature representation will better highlight the dynamic trends of device states, helping to improve the accuracy of subsequent fault diagnosis, specifically as follows:

[0087] For the adaptive dynamic pyramid Layer Aggregate features , define its fusion weight is the normalized exponential function form: ,in, , is a hyperparameter that controls the degree of concentration of fusion weight distribution, This is the first Tier Aggregate features at time Volatility indicator.

[0088] Intuitively, based on the fusion weights, the volatility index of the aggregated features is mapped to The interval can be processed by Softmax normalization so that the sum of the weights of all aggregated features in the same layer is exactly 1. Hyperparameters Controls the concentration of fusion weight distribution: The larger the value, the more the fusion weight distribution is concentrated on a few features with higher volatility indicators; on the contrary, The smaller is, the more uniform the fusion weight distribution is. In practical application, Adjustments can be made based on historical data and experience in order to achieve the optimal balance of fusion weight distribution.

[0089] Based on the fusion weight, layer Aggregate features Perform weighted fusion to obtain Layer fusion features : ,in, for dimensional column vector, which comprehensively reflects the dynamic information of each aggregate feature at the ℓth layer time granularity, and the aggregate feature with higher volatility index contributes more to the fusion result. The larger the proportion.

[0090] Furthermore, the adaptive dynamic pyramid Fusion features of each layer Perform weighted fusion in the inter-layer direction to form a global device state feature vector : ,in, For the The global weight of the layer fusion feature satisfies the normalization condition , Indicates the total number of layers of the adaptive dynamic pyramid.

[0091] For example, the global weight setting strategy may be: Set as Average volatility index of the layer The normalized value of , that is: ,in is the layer number index variable, and its value range is from the 1st layer to the layer, As the iterative variable of the summation, it ensures that the sum of the weights of all layers is 1; the pyramid layer with a larger average volatility index value can occupy a larger global weight in weighted fusion, making It places more emphasis on time scales with strong representation capabilities. The aggregated features of this layer change more dramatically overall, that is, it contains more meaningful information about dynamic changes in devices.

[0092] In summary, by using the adaptive dynamic pyramid Based on the introduction of adaptive feature weighted fusion mechanism, a global fusion device state feature vector is obtained. Compared with the original multi-source heterogeneous data, The information representation is more refined and enriched, and can dynamically depict the health status of the equipment from both time and space dimensions. This lays a solid feature foundation for subsequent fault diagnosis and prediction.

[0093] It should be noted that the feature fusion framework described in this step has a certain degree of universality and can be appropriately simplified or improved according to actual needs. For example, if the real-time diagnosis is required to be high, the global weighted fusion between pyramid layers can be omitted and the weighted fusion of each layer can be directly combined. Input diagnostic module in parallel; if you want to further explore the correlation between inter-layer features, you can A sequence dependency modeling unit similar to the Long Short-Term Memory (LSTM) is introduced to characterize the dynamic dependencies between pyramids at different resolutions. These are all extensions of the adaptive feature fusion framework of the present invention in specific applications.

[0094] Step 5: Fault diagnosis optimization based on multi-armed bandit

[0095] It should be noted that the dynamic pyramid model and adaptive feature fusion mechanism constructed in the aforementioned steps provide a rich, dynamic feature representation for smart device fault diagnosis. However, making accurate and reliable fault determinations based on these features and continuously optimizing diagnostic strategies to adapt to the dynamic evolution of device status remain key challenges. To this end, this paper further introduces a diagnostic strategy optimization mechanism based on a multi-armed bandit (MAB). This mechanism dynamically optimizes diagnostic strategies to adapt to the evolving device status, maximizing diagnostic accuracy and reliability while minimizing long-term diagnostic risk.

[0096] MAB is a classic problem in the field of reinforcement learning. Its core concept is to seek the optimal balance between exploration and exploitation, finding the decision with the highest expected benefit through continuous trial and learning. This paper introduces this into the fault diagnosis scenario to achieve adaptive optimization of diagnostic results: by continuously exploring the diagnostic effects of different fault modes, the diagnostic strategy is dynamically adjusted to adapt to changes in device status and maximize the overall diagnostic benefit. The specific implementation of the diagnostic strategy optimization mechanism based on the multi-armed bandit (MAB) includes:

[0097] (1) Fault feature dictionary construction:

[0098] Offline builds include Feature dictionary of known equipment failure modes, denoted as , where each , the feature vector corresponding to a certain failure mode can be obtained by analyzing historical failure data and summarizing expert experience. The dimension of the feature vector and the device state feature vector in step 4 The dimensions are the same.

[0099] (2) Online fault diagnosis and similarity matching:

[0100] Perform online diagnosis on the equipment at each time step t and convert the equipment state feature vector With feature dictionary The characteristic vector of each failure mode in Compare and calculate the similarity between the two :

[0101]

[0102] in, It is a similarity measurement function, and its specific forms include Euclidean distance, cosine similarity, and Pearson correlation coefficient.

[0103] Based on the similarity , the equipment fault diagnosis problem is formalized as a multi-armed bandit problem, where each fault mode m corresponds to a bandit arm, and its expected payoff is It is defined as the negative similarity score as follows:

[0104]

[0105] Smaller (i.e., higher similarity) indicates that the fault mode m is similar to the current device state feature vector The higher the likelihood of a match, the lower the corresponding diagnostic risk.

[0106] (3) Adaptive strategy optimization based on Thompson sampling algorithm:

[0107] To minimize the long-term cumulative diagnostic risk, Select a range of diagnostic results , i.e. the failure mode, so that the total diagnostic risk Minimum: The present invention uses a strategy selection algorithm based on Thompson sampling. This algorithm achieves a dynamic balance between exploration (trying failure modes with higher uncertainty) and exploitation (selecting failure modes with higher confidence) to optimize the diagnostic strategy. The specific steps are as follows:

[0108] a. Initialization:

[0109] For each failure mode m, a parameter is maintained as and Beta distribution , which represents its expected return Initially, set , , which is applicable to all m.

[0110] b. Sampling and selection:

[0111] At each time step t, for each failure mode m, from its Beta distribution Sampling an estimated return value : ;

[0112] Select the failure mode with the largest estimated benefit : ;

[0113] Output failure mode The corresponding eigenvector , as the current diagnosis result.

[0114] c. Feedback and Updates:

[0115] Observe the true state of the device at time step t+1 ,and ,in Indicates that failure mode m occurs, Indicates a diagnostic error or the device is normal. Update the Beta distribution parameters as follows:

[0116] like (the current diagnosis result is correct), then increase ,make: ;

[0117] like (The current diagnosis result is wrong), then increase ,make: ;

[0118] d. Iteration:

[0119] Let t increase, that is Repeat steps b to c until the diagnostic process is complete.

[0120] Through continuous parameter updates, the algorithm optimizes the posterior distribution, increases the probability of selecting fault modes with higher historical diagnostic accuracy, and explores fault modes with higher uncertainty to adapt to changes in equipment status.

[0121] (4) Diagnostic confidence assessment:

[0122] In order to improve the reliability of the diagnosis results, the present invention introduces the confidence index , , defined as the current device state feature vector The feature vector corresponding to the diagnosis result The normalized similarity of:

[0123]

[0124] When it is close to 1, it means that the current diagnosis result matches the actual status of the device more closely, and the diagnosis confidence is high; otherwise, When it is smaller, it indicates that the diagnostic uncertainty is high and further verification or maintenance measures are needed.

[0125] It should be noted that the fault diagnosis optimization mechanism based on the multi-armed bandit has the following advantages: Adaptability: Through the Thompson sampling algorithm, the diagnosis strategy is dynamically adjusted based on real-time equipment status feedback to adapt to changes in complex working conditions; High precision: By prioritizing high-similarity fault modes and optimizing the selection through iterative learning, the diagnostic accuracy is improved; Explanability: The confidence index Provide quantitative reliability evaluation for diagnostic results and provide important reference for equipment maintenance decisions.

[0126] In summary, this step fully leverages the dynamic feature representation generated by adaptive dynamic pyramids and feature fusion to achieve accurate, adaptive, and reliable fault diagnosis. This mechanism is universal and can be adjusted according to application requirements, such as by introducing additional similarity metrics or optimizing the exploration-exploitation balance through hyperparameter optimization.

[0127] This embodiment, through a series of innovative technical features, significantly overcomes the shortcomings of existing technologies in multi-source heterogeneous data processing, dynamic adaptability, and online optimization. Specifically, this embodiment first collects and preprocesses multi-source heterogeneous data (step 1), cleans, normalizes, and smoothes sensor data and operation logs, generating a standardized time series feature set that lays the foundation for subsequent analysis. Next, a dynamic pyramid model is constructed (step 2), aggregating features at multiple levels of time granularity. An adaptive structural adjustment mechanism based on a volatility index (coefficient of variation) (step 3) dynamically optimizes the number of layers and granularity to accurately capture the multi-scale dynamic characteristics of device status. Furthermore, this embodiment introduces an adaptive feature weighted fusion mechanism (step 4), dynamically assigning weights based on volatility indicators to generate a refined global feature vector. Finally, a multi-armed bandit-based fault diagnosis optimization mechanism (step 5) is employed, utilizing the Thompson sampling algorithm and the confidence index c(t) to achieve dynamic fault mode selection and reliability assessment. Through the above-mentioned technical features, this embodiment achieves efficient enrichment of feature representation, robust adaptation of the model to complex working conditions, and continuous optimization of diagnostic strategies, significantly improving the accuracy, reliability, and long-term performance of fault detection. It is particularly suitable for complex equipment management needs in industrial Internet and intelligent manufacturing scenarios.

[0128] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed herein. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present invention are indicated by the claims. It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and that various modifications and changes can be made without departing from its scope. The scope of the present invention is limited only by the appended claims.

Claims

1. A device fault detection method based on dynamic pyramid, characterized in that: include: Collect and preprocess multi-source heterogeneous data from devices to obtain a multi-source time series feature set, perform aggregation operations on the multi-source time series feature set at L different time granularities, obtain L layers of aggregated features, and construct an initial dynamic pyramid model; Based on the characteristic volatility index, adaptively adjusting the number of layers and time granularity of the dynamic pyramid model to generate an adaptive dynamic pyramid; The adaptive adjustment requires first adjusting each layer of the dynamic pyramid model , calculate the arithmetic mean of the characteristic volatility index of all its aggregated features as the average volatility intensity, and the calculation formula is as follows: ,in For each layer The number of aggregated features in , is the timestamp, is the characteristic volatility indicator, is the average volatility intensity; Based on the aggregated features of the adaptive dynamic pyramid, an adaptive feature weighted fusion mechanism is adopted to generate a global device state feature vector; Based on the multi-armed bandit machine, the fault diagnosis optimization of the device state feature vector is performed through the Thompson sampling algorithm, and the failure mode and confidence index of the device are output.

2. The device fault detection method based on dynamic pyramid according to claim 1, characterized in that: The initial dynamic pyramid model construction includes: Based on the time series feature set, a The dynamic pyramid model of the layers The layer time granularity is , is the sampling interval, , the time granularity increases from the bottom layer to the top layer; Perform aggregation operations on each time series feature at each layer to generate the first Tier Aggregate features , aggregate each feature exist The results of aggregation at each granularity level constitute the initial dynamic pyramid model : .

3. The device fault detection method based on dynamic pyramid according to claim 2, characterized in that: For the Pyramid Layer Aggregate features , its volatility index Defined as the aggregate feature in the historical time window coefficient of variation within ; The adaptive adjustment is based on the minimum preset threshold after calculating the average volatility intensity. and the highest preset threshold , determine the average volatility index of each layer Do they all fall within the interval? If not, adjust the number of layers and time granularity through splitting or merging operations until the average volatility indicators of all layers fall within the interval , and obtain the adaptive dynamic pyramid.

4. The device fault detection method based on dynamic pyramid according to claim 3, characterized in that: The splitting operation includes: When Layer average volatility intensity , in Layer and A split aggregation layer is inserted between the layers, and its time granularity is Layer and Arithmetic mean of layer time granularity: , and based on the time granularity, the aggregation characteristics of the split aggregation layer are generated as follows: ; The merging operation includes: Jordi Layer average volatility intensity , and the average volatility of the previous layer Close, that is: ,in is a small positive threshold, then Layer and Layer merging, generating time granularity is The combined aggregation layer of , based on time granularity Regenerate aggregate features of merged aggregate layers .

5. The device fault detection method based on dynamic pyramid according to claim 1, characterized in that: The adaptive feature weighted fusion includes: Based on the Layer Aggregate features Characteristic volatility indicator , define its fusion weight as: , in, , is a hyperparameter that controls the degree of concentration of fusion weight distribution; Based on the fusion weight, layer Aggregate features are weighted fused to generate the Layer fusion features: , and generate the global device state feature vector through global weighted fusion ,in For the The global weight of the layer fusion features, Indicates the total number of layers of the adaptive dynamic pyramid.

6. The device fault detection method based on dynamic pyramid according to claim 5, characterized in that: The global weight is set to Layer Average Volatility Index The normalized value of , that is: ,in is the layer number index variable, and its value range is from the 1st layer to the layers, so that the layers with higher volatility occupy a greater weight in the global fusion, and enhance the ability to represent dynamic changes.

7. The device fault detection method based on dynamic pyramid according to claim 1, characterized in that: The fault diagnosis optimization includes: offline construction including Feature dictionary of known equipment failure modes , where each The characteristic vector corresponding to a certain fault mode and the characteristic vector of the equipment state Same dimension; Online calculation of device state feature vectors Characteristics of each failure mode Similarity: ; Define expected return as: , the equipment fault diagnosis problem is formalized as a multi-armed bandit problem.

8. The device fault detection method based on dynamic pyramid according to claim 7, characterized in that: The Thompson sampling algorithm includes: For each failure mode The maintenance parameters are and Beta distribution , initial settings , ; At time step , estimated returns from various Beta distributions : , select the failure mode with the largest estimated benefit value , , and the failure mode The corresponding eigenvector , as the current diagnosis result; Update parameters according to the actual device status: If the current diagnosis result is correct, increase ,make: , otherwise increase ,make: , iteratively optimize the diagnostic strategy.

9. The device fault detection method based on dynamic pyramid according to claim 7, characterized in that: The confidence index is defined as: , used to evaluate the reliability of the current diagnostic results, A value close to 1 indicates that the current diagnosis result closely matches the device status.

10. The device fault detection method based on dynamic pyramid according to claim 7, characterized in that: The similarity measurement method includes Euclidean distance, cosine similarity or Pearson correlation coefficient, and the multi-source heterogeneous data includes sensor data, equipment operation logs and environmental parameters.

Images