Multi-agent-based gearbox real-time monitoring and intelligent early warning system

Abstract

The invention discloses a gearbox real-time monitoring and intelligent early warning system based on multiple agents, and the system comprises a signal collection module which is used for collecting vibration, acoustic emission and rotating speed signals and generating a monitoring sequence; the time sequence construction module is used for time delay embedding and forming a high-dimensional vector set; the anomaly modeling module is used for constructing an RRCF model and outputting a deviation score sequence; the level detection module is used for generating short-time abnormity and long-term degradation results; the distribution modeling module is used for establishing a risk distribution model and forming a dynamic threshold value; the structure updating module is used for adjusting the tree structure and updating the model during insertion, deletion and working condition change; and the consensus aggregation module is used for running the RRCF in the intelligent body, generating a consistent risk index based on consensus, and outputting early warning information when the consistent risk index exceeds a threshold value. According to the invention, through the cooperation of the improved RRCF and multiple agents, the real-time monitoring and early warning of the fault of the gearbox are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and equipment condition monitoring technology, and in particular to a real-time monitoring and intelligent early warning system for gearboxes based on multi-agent systems. Background Technology

[0002] As a core component of mechanical transmission systems, the gearbox's operating status directly affects the stability and reliability of the entire machine. In industrial production, energy equipment, transportation, and intelligent manufacturing, gearboxes often operate under high loads and complex conditions, making them prone to early failures such as wear, cracks, and pitting. Existing gearbox monitoring technologies mainly rely on vibration analysis, acoustic emission detection, and speed measurement. These methods acquire signals through sensors and combine them with algorithms such as spectrum analysis, wavelet transform, or envelope demodulation to identify fault features. However, these methods generally suffer from drawbacks such as strong reliance on prior knowledge, limited ability to process non-stationary signals, and reliance on human experience for feature extraction. They are also ill-suited to handling dynamic changes under complex operating conditions and the strong coupling characteristics of multi-source signals.

[0003] In recent years, with the development of machine learning and anomaly detection algorithms, some techniques based on statistical modeling and learning methods have been gradually applied to equipment condition monitoring. For example, methods such as Isolation Forest, local outlier factor, and principal component analysis have been used for anomaly detection. However, when processing high-dimensional time-series signals, they are prone to problems such as insufficient feature dimension selection, high noise sensitivity, and poor detection stability. Especially in scenarios like gearboxes, which involve the coupling of vibration signals, acoustic emission signals, and rotational speed signals, existing methods cannot fully utilize the topological correlation between the signals, resulting in insufficient recognition rate of abnormal patterns and difficulty in timely detection of early faults.

[0004] Existing research on randomly partitioned forests has certain advantages in anomaly detection, as it can measure the isolation degree of samples by randomly partitioning the space, thereby identifying potential anomalies. However, traditional RRCF models lack constraints on the selection of partitioning dimensions and the determination of partitioning points, failing to incorporate the structural topological relationships between gears, drive shafts, and bearings, resulting in partitioning results lacking physical meaning. Furthermore, conventional randomly partitioned forests are prone to imbalanced partitioning when processing non-uniformly distributed high-dimensional time-series signals, leading to distorted bias score calculations and a lack of stability and reliability in anomaly detection results.

[0005] Furthermore, most existing anomaly detection methods only output results at a single time scale, lacking hierarchical modeling for both short-term anomalies and long-term degradation, and thus failing to simultaneously reflect the risks of sudden failures and gradual degradation. In terms of risk modeling, traditional methods typically use fixed thresholds to determine equipment status, which is ill-suited to the dynamic changes in signal distribution under complex operating conditions, easily leading to missed or false alarms. Regarding model updates, existing methods are prone to tree structure imbalance under frequent insertions and deletions of high-dimensional time-series signals, lacking adaptive adjustment mechanisms and making it difficult to maintain the long-term effectiveness of the model.

[0006] In distributed application scenarios, existing methods often rely on single-point modeling and analysis, lacking coordination and consensus mechanisms among multiple agents. This makes it impossible to achieve unified risk assessment in large-scale equipment group monitoring and smart factory scenarios. Even when some studies introduce multi-agent frameworks, they mostly remain at the level of data collection and simple result fusion, lacking mechanisms to generate a global risk index through consensus calculation, making it difficult to support the high-reliability intelligent early warning requirements.

[0007] Therefore, how to provide a real-time monitoring and intelligent early warning system for gearboxes based on multi-agent systems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a real-time monitoring and intelligent early warning system for gearboxes based on multi-agent systems. This invention utilizes vibration signals, acoustic emission signals, and rotational speed signals to construct a high-dimensional time-series vector set. It combines this with an RRCF model incorporating topological constraints and quantile density cutting to calculate deviation scores. Based on a hierarchical structure, it simultaneously generates short-term anomaly detection results and long-term degradation detection results. Furthermore, it establishes a dynamic threshold set through risk distribution modeling, adaptively adjusting the tree structure and maintaining model stability during signal insertion and deletion. Finally, it generates a consistent risk index and outputs gearbox early warning information through multi-agent collaboration and consensus algorithms. This system boasts advantages such as high detection accuracy, strong adaptability to complex operating conditions, sensitivity to early faults, timely model updates, and support for intelligent monitoring of grouped equipment.

[0009] A real-time monitoring and intelligent early warning system for gearboxes based on a multi-agent system according to an embodiment of the present invention includes:

[0010] The signal acquisition module is used to acquire vibration signals, acoustic emission signals, and rotational speed signals and generate standardized monitoring sequences.

[0011] The time series construction module is used to perform time delay embedding on standardized monitoring sequences, generate high-dimensional time series vectors and establish timestamp indexes to form a high-dimensional time series vector set.

[0012] The anomaly modeling module is used to build an RRCF model on a high-dimensional time-series vector set. It selects the cutting dimension based on topological constraints and determines the cutting point based on quantile density. It records the path length when inserting vectors and calculates the expected path length based on the number of leaf node samples. It determines the deviation score based on the difference and generates a deviation score sequence by averaging in the forest.

[0013] The hierarchical detection module is used to generate short-term anomaly detection results in the hierarchical RRCF model using a sliding window, and to generate long-term degradation detection results within the aggregation period.

[0014] The distribution modeling module is used to input short-term anomaly detection results and long-term degradation detection results into the risk distribution model, calculate statistics and establish a distribution function, determine dynamic thresholds at quantiles and form a dynamic threshold set.

[0015] The structure update module is used to update the tree structure when inserting or deleting high-dimensional time-series vectors, adjust the number and depth of trees when the rate of change of speed signal and load signal exceeds the threshold, perform subtree reconstruction when the tree structure is unbalanced, and recalculate the deviation score and update the risk distribution model and dynamic threshold set after the update.

[0016] The consensus aggregation module is used to run a hierarchical RRCF model in each intelligent entity to calculate the local risk parameter set. The consensus module generates a consistent risk index, compares it with a dynamic threshold set, and outputs a gearbox warning message and records a timestamp when the index exceeds the threshold.

[0017] Optionally, modules can be integrated using the following methods:

[0018] The gearbox vibration signal, acoustic emission signal and speed signal are collected, and time synchronization and normalization processing is performed to obtain a standardized monitoring sequence.

[0019] Time delay embedding is performed on standardized monitoring sequences to form a high-dimensional time series vector set;

[0020] An RRCF model is constructed on a high-dimensional time-series vector set. Topological relationship constraints of gears, drive shafts and bearings are introduced during the selection of cutting dimensions. Non-uniform cutting is performed based on quantile density, and the deviation fraction sequence is output.

[0021] The RRCF model is set as a hierarchical structure. The first layer processes the deviation score sequence and outputs short-term anomaly detection results. The second layer aggregates the short-term anomaly detection results within the time window and outputs long-term degradation detection results.

[0022] Input the short-term anomaly detection results and long-term degradation detection results into the distribution modeling module to establish a risk distribution model and calculate a dynamic threshold set;

[0023] When inserting and deleting high-dimensional time-series vectors, the number and depth of trees are adjusted according to the rotation speed and load signals. When the structure is unbalanced, node updates and subtree reconstruction are performed, and the risk distribution model and dynamic threshold set are updated simultaneously.

[0024] Each agent independently runs a hierarchical RRCF model and outputs local risk parameters. All agents aggregate local risk parameters based on a consensus algorithm to generate a consistent risk index, and output gearbox warning information based on the consistent risk index and a dynamic threshold set.

[0025] Optionally, the generation of the high-dimensional time-series vector set specifically includes:

[0026] Set the embedding dimension to a positive integer, the delay time to a sample interval of not less than one, the window length to the product of the embedding dimension and the delay time, and the step size to a fixed number of samples.

[0027] Sliding truncation is performed on the standardized monitoring sequence according to the window length and step size. At the beginning of each window, observations of the number of embedded dimensions are extracted sequentially at fixed intervals of delay time to form a delayed sampling sequence.

[0028] Delay sampling is performed on vibration signal, acoustic emission signal and rotation speed signal respectively. The three types of delay sampling results are spliced ​​in the order of vibration signal, acoustic emission signal and rotation speed signal. Within each type of signal, they are arranged in the order from zero delay to maximum delay to generate a single high-dimensional time vector.

[0029] The current window is discarded when the starting point of the window cannot cover the complete number of embedded dimensions.

[0030] For each high-dimensional time series vector, record the start timestamp, end timestamp, channel order, and delay order, and establish the correspondence between vector index and timestamp;

[0031] All high-dimensional time series vectors are combined in chronological order to form a high-dimensional time series vector set, which is used to construct the input of the RRCF model.

[0032] Optionally, the construction and processing of the RRCF model specifically includes:

[0033] An RRCF model is constructed on a high-dimensional time series vector set, with the number of trees set to a fixed integer and the maximum depth of each tree set to a fixed integer.

[0034] When selecting the cutting dimension, the topological relationship between gears, drive shafts and bearings is represented as an adjacency matrix. Based on the adjacency matrix, the feature dimensions with connection relationships are divided into candidate sets, and a dimension is randomly selected from the candidate set as the cutting dimension.

[0035] When determining the cut point, the range of values ​​of the selected cut dimension is used as the cut object. The range of values ​​is divided according to the quantiles. The interval density is calculated based on the number of observations in each quantile interval. The interval density is converted into sampling weights. The interval is selected according to the sampling weights. The cut point is randomly determined within the interval.

[0036] After determining the cutting dimension and cutting point, the space where the high-dimensional time series vector is located is divided into two subspaces. Based on the relationship between the value of the high-dimensional time series vector in the cutting dimension and the size of the cutting point, the high-dimensional time series vector is assigned to the corresponding subspace.

[0037] During the recursive partitioning process, the selection of cutting dimensions and determination of cutting points are repeatedly executed until the maximum depth of the tree or the capacity limit of the leaf nodes is reached.

[0038] When inserting vectors into the tree structure, record the cutting path length and cutting depth, and determine the deviation score based on the difference between the cutting path length and the expected path length calculated based on the number of leaf node samples.

[0039] After the insertion operation is completed in all trees, the deviation scores of the same high-dimensional time series vector in all trees are averaged to generate a deviation score sequence.

[0040] Optionally, the hierarchical processing of the RRCF model specifically includes:

[0041] In the hierarchical RRCF model, define the first and second layers;

[0042] In the first layer, the sliding window length is set to a fixed integer. The deviation score sequence is divided according to the sliding window order. The arithmetic mean of all deviation scores is calculated in each sliding window. The variance of all deviation scores is calculated. The maximum value of all deviation scores is calculated. The arithmetic mean, variance and maximum value are combined to form the short-term anomaly detection result.

[0043] In the second layer, the short-term anomaly detection results are arranged in chronological order. The aggregation period is set to a fixed integer. The short-term anomaly detection results within one aggregation period are used as input. Within this input, the average of all arithmetic means, the average of all variances, and the maximum of all maximum values ​​are calculated. The average and maximum results are combined to form the long-term degradation detection results.

[0044] Optionally, establishing the risk distribution model and calculating the dynamic threshold set specifically includes:

[0045] Input the short-term anomaly detection results and long-term degradation detection results into the distribution modeling module;

[0046] In the distributed modeling module, short-term anomaly detection results and long-term degradation detection results are collected within a fixed time range to form a risk observation sample set;

[0047] Calculate the mean, variance, and quantiles of the risk observation sample set, and use the mean and variance as the initial statistics for distribution modeling;

[0048] When establishing a risk distribution model, the relationship between the probability density function, cumulative distribution function, and dynamic threshold is defined as follows:

[0049] ;

[0050] in, Indicates the risk observation value. Represents a set of parameters. The symbol represents the energy function, and the integral symbol indicates a normalized calculation over the entire space. Represents the probability density function. Represents the cumulative distribution function. Indicates quantiles, This represents the dynamic threshold corresponding to the quantile. It represents the inverse function of the cumulative distribution function;

[0051] All dynamic thresholds are combined into a dynamic threshold set, a timestamp is recorded for each dynamic threshold, and a correspondence is established with the risk observation sample set.

[0052] Optionally, the updating of the risk distribution model and the dynamic threshold set specifically includes:

[0053] When a high-dimensional time series vector is inserted into the set, the newly added high-dimensional time series vector is sequentially input into all tree structures. During the input process, the path nodes are determined according to the cutting dimension and cutting point, and the timestamp of the newly added high-dimensional time series vector is registered at the leaf node position.

[0054] When a deletion occurs in the high-dimensional time-series vector set, the timestamp of the deleted high-dimensional time-series vector is used to locate the leaf node in the tree structure, the current leaf node record is removed, and the associated cutting path is updated.

[0055] When the speed signal and load signal change, calculate the rate of change and compare it with a preset threshold. When the rate of change is greater than the threshold, increase the number of trees or increase the depth of the trees. When the rate of change is less than the threshold, decrease the number of trees or decrease the depth of the trees.

[0056] After the number of trees or the depth of the trees are adjusted, the balance of the tree structure is checked. When the tree structure is unbalanced, subtree reconstruction is performed on the affected branches, the nodes are re-divided into subspaces, and the cutting dimensions and cutting points are reallocated to the subspaces.

[0057] After completing the insertion, deletion and tree structure adjustment of high-dimensional time series vectors, the deviation score is recalculated and the new deviation score is input into the risk distribution model;

[0058] After the risk distribution model is updated, the dynamic threshold set is recalculated, a timestamp is recorded for each dynamic threshold, and a correspondence between the dynamic threshold and the risk observation sample set is established.

[0059] Optionally, the generation of the warning information specifically includes:

[0060] Run a hierarchical RRCF model in each intelligent body, input the local bias score sequence, and obtain the local short-term anomaly detection results, local long-term degradation detection results, and local risk distribution model;

[0061] In each intelligent entity, a local risk parameter set is calculated based on the local risk distribution model, and the local risk parameter set is transmitted to the consensus module;

[0062] In the consensus module, the local risk parameter set transmitted by all intelligent agents is received, consensus calculation is performed, and the consensus risk index is output.

[0063] The consistency risk index is compared with a dynamic threshold set. When the consistency risk index is greater than the threshold in the dynamic threshold set, a gearbox warning message is generated.

[0064] When generating gearbox warning information, record the consistency risk index value, the triggered dynamic threshold and the corresponding timestamp, and output the warning information.

[0065] The beneficial effects of this invention are:

[0066] This invention overcomes the problems of strong randomness in the selection of cutting dimensions and lack of physical meaning in the cutting results of traditional methods by constructing an RRCF model on a high-dimensional time-series vector set, which introduces topological constraints and quantile density cutting. It achieves constraint modeling based on the structural relationships of gears, drive shafts, and bearings, thus more accurately reflecting the coupling characteristics between signals. In the deviation score calculation process, the difference between the path length and the expected path length is used to measure the degree of anomaly, and the average is taken at the forest level, improving the stability and reliability of anomaly detection results.

[0067] This invention sets the RRCF model as a hierarchical structure, which can generate both short-term anomaly detection results through a sliding window and long-term degradation detection results within an aggregation period, thus solving the problem that traditional methods cannot simultaneously handle sudden faults and gradual degradation. In terms of risk assessment, this invention introduces a distributed modeling mechanism, making judgments based on a dynamic threshold set. This overcomes the limitation of fixed threshold models, which are prone to false alarms and false negatives under complex operating conditions, and enhances the system's adaptability to non-stationary signals.

[0068] This invention further incorporates an adaptive tree structure adjustment mechanism for high-dimensional time-series vector insertion and deletion. This mechanism automatically adjusts the number and depth of trees based on changes in rotational speed and load signals, and performs subtree reconstruction when structural imbalance occurs, thereby ensuring long-term stable model operation. Simultaneously, this invention utilizes independent operation of multiple agents and consensus aggregation to generate a consistent risk index and output early warning information by combining it with a dynamic threshold set. This achieves unified risk assessment in distributed equipment monitoring scenarios, enhancing the capabilities of grouped equipment management and intelligent early warning. Attached Figure Description

[0069] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0070] Figure 1 This is a flowchart of a multi-agent-based real-time monitoring and intelligent early warning system for gearboxes proposed in this invention;

[0071] Figure 2 This is a schematic diagram of the RRCF model structure of a multi-agent-based gearbox real-time monitoring and intelligent early warning system proposed in this invention; Figure 3 This is a schematic diagram of multi-agent consensus aggregation for a real-time monitoring and intelligent early warning system for gearboxes based on multi-agents proposed in this invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0073] refer to Figure 1-3 A real-time monitoring and intelligent early warning system for gearboxes based on multi-agent systems, comprising:

[0074] The signal acquisition module is used to acquire vibration signals, acoustic emission signals, and rotational speed signals and generate standardized monitoring sequences.

[0075] The time series construction module is used to perform time delay embedding on standardized monitoring sequences, generate high-dimensional time series vectors and establish timestamp indexes to form a high-dimensional time series vector set.

[0076] The anomaly modeling module is used to build an RRCF model on a high-dimensional time-series vector set. It selects the cutting dimension based on topological constraints and determines the cutting point based on quantile density. It records the path length when inserting vectors and calculates the expected path length based on the number of leaf node samples. It determines the deviation score based on the difference and generates a deviation score sequence by averaging in the forest.

[0077] The hierarchical detection module is used to generate short-term anomaly detection results in the hierarchical RRCF model using a sliding window, and to generate long-term degradation detection results within the aggregation period.

[0078] The distribution modeling module is used to input short-term anomaly detection results and long-term degradation detection results into the risk distribution model, calculate statistics and establish a distribution function, determine dynamic thresholds at quantiles and form a dynamic threshold set.

[0079] The structure update module is used to update the tree structure when inserting or deleting high-dimensional time-series vectors, adjust the number and depth of trees when the rate of change of speed signal and load signal exceeds the threshold, perform subtree reconstruction when the tree structure is unbalanced, and recalculate the deviation score and update the risk distribution model and dynamic threshold set after the update.

[0080] The consensus aggregation module is used to run a hierarchical RRCF model in each intelligent entity to calculate the local risk parameter set. The consensus module generates a consistent risk index, compares it with a dynamic threshold set, and outputs a gearbox warning message and records a timestamp when the index exceeds the threshold.

[0081] This invention organically combines a signal acquisition module, a timing construction module, an anomaly modeling module, a hierarchical detection module, a distributed modeling module, a structure update module, and a consensus aggregation module to form a complete real-time monitoring and intelligent early warning system for gearboxes. Based on data acquisition, feature construction, anomaly modeling, hierarchical detection, risk distribution modeling, and adaptive structure updating, a multi-agent consensus mechanism is introduced to achieve result aggregation, effectively avoiding the bias problem of single-point detection. This ensures that accurate early warning information can be stably output under complex operating conditions, making gearbox fault detection more timely and risk assessment more reliable.

[0082] In this embodiment, the modules are interconnected using the following method:

[0083] The gearbox vibration signal, acoustic emission signal and speed signal are collected, and time synchronization and normalization processing is performed to obtain a standardized monitoring sequence.

[0084] Time delay embedding is performed on standardized monitoring sequences to form a high-dimensional time series vector set;

[0085] An RRCF model is constructed on a high-dimensional time-series vector set. Topological relationship constraints of gears, drive shafts and bearings are introduced during the selection of cutting dimensions. Non-uniform cutting is performed based on quantile density, and the deviation fraction sequence is output.

[0086] The RRCF model is set as a hierarchical structure. The first layer processes the deviation score sequence and outputs short-term anomaly detection results. The second layer aggregates the short-term anomaly detection results within the time window and outputs long-term degradation detection results.

[0087] Input the short-term anomaly detection results and long-term degradation detection results into the distribution modeling module to establish a risk distribution model and calculate a dynamic threshold set;

[0088] When inserting and deleting high-dimensional time-series vectors, the number and depth of trees are adjusted according to the rotation speed and load signals. When the structure is unbalanced, node updates and subtree reconstruction are performed, and the risk distribution model and dynamic threshold set are updated simultaneously.

[0089] Each agent independently runs a hierarchical RRCF model and outputs local risk parameters. All agents aggregate local risk parameters based on a consensus algorithm to generate a consistent risk index, and output gearbox warning information based on the consistent risk index and a dynamic threshold set.

[0090] This invention generates standardized monitoring sequences by performing time synchronization and normalization processing on the collected monitoring signals. It then transforms a single sequence into a high-dimensional time-series vector set through time delay embedding. This solves the problems of insufficient dimensionality of the original signals and the inability of single-point features to fully reflect the gearbox operating status, ensuring the integrity of the time-series features required for subsequent modeling. At the same time, it establishes a strict time-series correspondence through timestamp indexing, enabling different signals to be analyzed under a unified coordinate system, thus laying a solid foundation for anomaly modeling and risk assessment.

[0091] In this embodiment, the generation of the high-dimensional time-series vector set specifically includes:

[0092] Set the embedding dimension to a positive integer, the delay time to a sample interval of not less than one, the window length to the product of the embedding dimension and the delay time, and the step size to a fixed number of samples.

[0093] Sliding truncation is performed on the standardized monitoring sequence according to the window length and step size. At the beginning of each window, observations of the number of embedded dimensions are extracted sequentially at fixed intervals of delay time to form a delayed sampling sequence.

[0094] Delay sampling is performed on vibration signal, acoustic emission signal and rotation speed signal respectively. The three types of delay sampling results are spliced ​​in the order of vibration signal, acoustic emission signal and rotation speed signal. Within each type of signal, they are arranged in the order from zero delay to maximum delay to generate a single high-dimensional time vector.

[0095] The current window is discarded when the starting point of the window cannot cover the complete number of embedded dimensions.

[0096] For each high-dimensional time series vector, record the start timestamp, end timestamp, channel order, and delay order, and establish the correspondence between vector index and timestamp;

[0097] All high-dimensional time series vectors are combined in chronological order to form a high-dimensional time series vector set, which is used to construct the input of the RRCF model.

[0098] This invention constructs an RRCF model on a high-dimensional time-series vector set, introduces topological relationship constraints of gears, drive shafts, and bearings in the selection of cutting dimensions, and determines the cutting point by combining quantile density. This avoids the randomness of cutting dimension selection in traditional methods and can better reflect the coupling relationship between mechanical structures. At the same time, the difference between the path length and the expected path length is introduced as a metric in the deviation score calculation, which improves the sensitivity to early anomalies. The final output deviation score sequence is more accurate and has a stronger ability to identify anomaly patterns.

[0099] In this embodiment, the construction and processing of the RRCF model specifically includes:

[0100] An RRCF model is constructed on a high-dimensional time series vector set, with the number of trees set to a fixed integer and the maximum depth of each tree set to a fixed integer.

[0101] When selecting the cutting dimension, the topological relationship between gears, drive shafts and bearings is represented as an adjacency matrix. Based on the adjacency matrix, the feature dimensions with connection relationships are divided into candidate sets, and a dimension is randomly selected from the candidate set as the cutting dimension.

[0102] When determining the cut point, the range of values ​​of the selected cut dimension is used as the cut object. The range of values ​​is divided according to the quantiles. The interval density is calculated based on the number of observations in each quantile interval. The interval density is converted into sampling weights. The interval is selected according to the sampling weights. The cut point is randomly determined within the interval.

[0103] After determining the cutting dimension and cutting point, the space where the high-dimensional time series vector is located is divided into two subspaces. Based on the relationship between the value of the high-dimensional time series vector in the cutting dimension and the size of the cutting point, the high-dimensional time series vector is assigned to the corresponding subspace.

[0104] During the recursive partitioning process, the selection of cutting dimensions and determination of cutting points are repeatedly executed until the maximum depth of the tree or the capacity limit of the leaf nodes is reached.

[0105] When inserting vectors into the tree structure, record the cutting path length and cutting depth, and determine the deviation score based on the difference between the cutting path length and the expected path length calculated based on the number of leaf node samples.

[0106] After the insertion operation is completed in all trees, the deviation scores of the same high-dimensional time series vector in all trees are averaged to generate a deviation score sequence.

[0107] This invention sets the RRCF model into a hierarchical structure, uses a sliding window to calculate the mean, variance, and maximum value of the deviation scores to form short-term anomaly detection results, and accumulates the short-term results within the aggregation period to generate long-term degradation detection results. This achieves a dual monitoring mechanism for short-term impact anomalies and long-term degradation trends. Compared with single-scale detection, this invention can capture diverse abnormal signals in gearbox operation at different time scales, thus reflecting the health status of the equipment more comprehensively.

[0108] In this embodiment, the hierarchical processing of the RRCF model specifically includes:

[0109] In the hierarchical RRCF model, define the first and second layers;

[0110] In the first layer, the sliding window length is set to a fixed integer. The deviation score sequence is divided according to the sliding window order. The arithmetic mean of all deviation scores is calculated in each sliding window. The variance of all deviation scores is calculated. The maximum value of all deviation scores is calculated. The arithmetic mean, variance and maximum value are combined to form the short-term anomaly detection result.

[0111] In the second layer, the short-term anomaly detection results are arranged in chronological order. The aggregation period is set to a fixed integer. The short-term anomaly detection results within one aggregation period are used as input. Within this input, the average of all arithmetic means, the average of all variances, and the maximum of all maximum values ​​are calculated. The average and maximum results are combined to form the long-term degradation detection results.

[0112] This invention inputs short-term anomaly detection results and long-term degradation detection results into a distribution modeling module, establishes a risk distribution model based on statistics, and determines a dynamic threshold set at quantiles. This allows the thresholds to be dynamically adjusted as the distribution of risk observation samples changes, avoiding false alarms and missed alarms caused by fixed thresholds under changing operating conditions. This improves the flexibility and accuracy of risk identification and ensures that the warning thresholds are consistent with the actual operating state of the gearbox.

[0113] In this embodiment, establishing the risk distribution model and calculating the dynamic threshold set specifically includes:

[0114] Input the short-term anomaly detection results and long-term degradation detection results into the distribution modeling module;

[0115] In the distributed modeling module, short-term anomaly detection results and long-term degradation detection results are collected within a fixed time range to form a risk observation sample set;

[0116] Calculate the mean, variance, and quantiles of the risk observation sample set, and use the mean and variance as the initial statistics for distribution modeling;

[0117] When establishing a risk distribution model, the relationship between the probability density function, cumulative distribution function, and dynamic threshold is defined as follows:

[0118] ;

[0119] in, Indicates the risk observation value. Represents a set of parameters. The symbol represents the energy function, and the integral symbol indicates a normalized calculation over the entire space. Represents the probability density function. Represents the cumulative distribution function. Indicates quantiles, This represents the dynamic threshold corresponding to the quantile. It represents the inverse function of the cumulative distribution function;

[0120] All dynamic thresholds are combined into a dynamic threshold set, a timestamp is recorded for each dynamic threshold, and a correspondence is established with the risk observation sample set.

[0121] This invention updates the tree structure during the insertion and deletion of high-dimensional time-series vectors, automatically adjusts the number and depth of trees when the rotational speed and load change rate exceed thresholds, and performs subtree reconstruction when the tree structure becomes unbalanced. This ensures the adaptability and stability of the model structure. After the structure adjustment is completed, the deviation score and risk distribution model are recalculated to ensure that the dynamic threshold set always matches the latest operating state, thereby significantly improving the robustness of the model under complex working conditions.

[0122] In this embodiment, the updating of the risk distribution model and the dynamic threshold set specifically includes:

[0123] When a high-dimensional time series vector is inserted into the set, the newly added high-dimensional time series vector is sequentially input into all tree structures. During the input process, the path nodes are determined according to the cutting dimension and cutting point, and the timestamp of the newly added high-dimensional time series vector is registered at the leaf node position.

[0124] When a deletion occurs in the high-dimensional time-series vector set, the timestamp of the deleted high-dimensional time-series vector is used to locate the leaf node in the tree structure, the current leaf node record is removed, and the associated cutting path is updated.

[0125] When the speed signal and load signal change, calculate the rate of change and compare it with a preset threshold. When the rate of change is greater than the threshold, increase the number of trees or increase the depth of the trees. When the rate of change is less than the threshold, decrease the number of trees or decrease the depth of the trees.

[0126] After the number of trees or the depth of the trees are adjusted, the balance of the tree structure is checked. When the tree structure is unbalanced, subtree reconstruction is performed on the affected branches, the nodes are re-divided into subspaces, and the cutting dimensions and cutting points are reallocated to the subspaces.

[0127] After completing the insertion, deletion and tree structure adjustment of high-dimensional time series vectors, the deviation score is recalculated and the new deviation score is input into the risk distribution model;

[0128] After the risk distribution model is updated, the dynamic threshold set is recalculated, a timestamp is recorded for each dynamic threshold, and a correspondence between the dynamic threshold and the risk observation sample set is established.

[0129] This invention calculates the local risk parameter set by independently running a hierarchical RRCF model in each intelligent body, and aggregates the results of all intelligent bodies in the consensus module to generate a consistent risk index. This effectively reduces the impact of local errors of a single intelligent body on the overall judgment, ensures the stability and consistency of the group monitoring results, and triggers an early warning and records a timestamp when the consistent risk index exceeds a dynamic threshold, thereby realizing highly reliable multi-device collaborative monitoring and early warning.

[0130] In this embodiment, the generation of the early warning information specifically includes:

[0131] Run a hierarchical RRCF model in each intelligent body, input the local bias score sequence, and obtain the local short-term anomaly detection results, local long-term degradation detection results, and local risk distribution model;

[0132] In each intelligent entity, a local risk parameter set is calculated based on the local risk distribution model, and the local risk parameter set is transmitted to the consensus module;

[0133] In the consensus module, the local risk parameter set transmitted by all intelligent agents is received, consensus calculation is performed, and the consensus risk index is output.

[0134] The consistency risk index is compared with a dynamic threshold set. When the consistency risk index is greater than the threshold in the dynamic threshold set, a gearbox warning message is generated.

[0135] When generating gearbox warning information, record the consistency risk index value, the triggered dynamic threshold and the corresponding timestamp, and output the warning information.

[0136] This invention compares a consistent risk index generated by a consensus aggregation module with a dynamic threshold set. When the risk index exceeds the threshold, it automatically generates and outputs an early warning message. At the same time, it records the risk index value, trigger threshold, and timestamp in detail, providing a reliable basis for anomaly analysis and fault tracing. Compared with the traditional single-point judgment mode, this invention ensures both the timeliness of early warning and data traceability, enabling the monitoring system to not only provide real-time early warnings but also provide accurate data support for subsequent maintenance and decision-making.

[0137] Example 1:

[0138] To verify the feasibility of this invention in practice, it was applied to a gearbox online monitoring and intelligent early warning scenario. The selected gearbox is a core component of industrial transmission equipment. Long-term operation under high speed and high load is prone to hidden dangers such as gear wear, bearing spalling, and transmission shaft imbalance. Traditional monitoring methods mainly rely on manual inspection and spectrum-based analysis methods, which are often slow to respond to early faults and are greatly affected by noise, resulting in low detection rate and high false alarm rate. This invention breaks through the limitations of traditional methods in terms of real-time performance and accuracy by combining a multi-agent architecture with an improved RRCF model.

[0139] During the application, sensors are placed at key locations in the gearbox. Vibration sensors collect vibration signals at a frequency of 10,000 points per second, acoustic emission sensors collect acoustic emission signals at a frequency of 20,000 points per second, and speed sensors collect speed signals at a frequency of 500 points per second. All signals are processed with a unified time reference and normalization. The generated standardized monitoring sequence is transformed into a high-dimensional time-series vector set through time delay embedding. Each vector has a timestamp index for subsequent modeling.

[0140] When constructing the RRCF model on a high-dimensional time-series vector set, this invention introduces the topological relationship constraints of gears, drive shafts, and bearings to define the cutting dimension, and determines the cutting point location through quantile density. The forest size is set to 100 trees, with a maximum depth of 50 for each tree. The deviation score is calculated by the difference between the actual path length and the expected path length and averaged in the forest to form a deviation score sequence. Experimental results show that under the initial gear wear signal, the deviation score increases by an average of about 35%, indicating that abnormal features are more prominent in the model of this invention.

[0141] In hierarchical detection, this invention achieves short-term and long-term anomaly detection through a two-layer structure. In the first layer, a 10-second sliding window is used to process the deviation score sequence and calculate the arithmetic mean, variance, and maximum value to form short-term anomaly detection results. In the second layer, the short-term results are accumulated and statistically analyzed with a 1-hour aggregation period to form long-term degradation detection results. In a 100-hour operation experiment, this invention can capture short-term impact signals within 10 hours of operation and gradually identify degradation trends during operation, while traditional methods failed to make effective identification within the same time range.

[0142] In risk assessment, this invention collects short-term and long-term detection results and establishes a risk distribution model, calculates the mean, variance, and quantiles, and determines a dynamic threshold set based on the quantiles. The dynamic thresholds can be adjusted according to the signal distribution, avoiding misjudgments caused by fixed thresholds under load changes. In comparative experiments, the traditional method only detected 7 out of 10 real anomalies, while this invention detected 10, increasing the detection rate from 65% to 92%, indicating that this invention is more sensitive in early fault identification.

[0143] In structural updates, this invention can adjust the number and depth of trees according to the rate of change of rotational speed and load, and perform subtree reconstruction when the structure is unbalanced, ensuring the continuous stability of the model. In a rapid load change experiment, the false alarm rate of the traditional method reached 12% within 30 minutes of operation, while the update mechanism of this invention reduced the false alarm rate to 4%, proving that this invention is more adaptable under complex dynamic working conditions.

[0144] In multi-agent collaborative applications, each agent independently runs a hierarchical RRCF model and generates a local risk parameter set. All agents aggregate through a consensus algorithm to form a consistent risk index. When the index exceeds a dynamic threshold set, an early warning message is output and a timestamp is recorded. In a scenario where 20 gearboxes are running simultaneously, this invention issues an early warning on average within 2 minutes of an anomaly occurring, while the traditional method has an average delay of more than 10 minutes. This result shows that the consensus mechanism significantly improves the overall response efficiency in multi-device scenarios.

[0145] The experimental data comparison is shown in Table 1:

[0146] Table 1 Comparison of Gearbox Monitoring and Early Warning Performance

[0147] index Traditional method (spectrum + fixed threshold) This invention (improved RRCF+ dynamic threshold) Early fault detection rate 65% 92% Average delay of short-term anomaly detection 8 minutes 2 minutes Accuracy of long-term degradation trend identification 70% 90% Anomaly detection false alarm rate 15% 5% Accuracy of risk assessment for multi-device consistency Unable to achieve 95%

[0148] A detailed comparison of the table reveals that the invention improves the early fault detection rate from 65% to 92%, demonstrating a significantly enhanced ability to capture minor anomalies. The average delay for short-term anomaly detection is reduced to 2 minutes, compared to 8 minutes for traditional methods, indicating superior real-time performance. The accuracy rate for identifying long-term degradation trends reaches 90%, while traditional methods only achieve 70%, demonstrating more reliable tracking of the degradation process. The false alarm rate for anomaly detection is 5%, a two-thirds reduction from the 15% of traditional methods, indicating stronger adaptability to complex operating conditions. Furthermore, the invention achieves 95% accuracy in multi-device consistent risk assessment, a feat unattainable by traditional methods, demonstrating good stability and accuracy in collaborative monitoring of group equipment. Therefore, data comparison confirms that the invention can achieve stable, real-time, and high-precision gearbox monitoring and intelligent early warning under complex operating conditions, validating its feasibility and effectiveness.

[0149] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A real-time monitoring and intelligent early warning system for gearboxes based on multi-agent technology, characterized in that, include: The signal acquisition module is used to acquire vibration signals, acoustic emission signals, and rotational speed signals and generate standardized monitoring sequences. The time series construction module is used to perform time delay embedding on standardized monitoring sequences, generate high-dimensional time series vectors and establish timestamp indexes to form a high-dimensional time series vector set. The anomaly modeling module is used to build an RRCF model on a high-dimensional time-series vector set. It selects the cutting dimension based on topological constraints and determines the cutting point based on quantile density. It records the path length when inserting vectors and calculates the expected path length based on the number of leaf node samples. It determines the deviation score based on the difference and generates a deviation score sequence by averaging in the forest. The hierarchical detection module is used to generate short-term anomaly detection results in the hierarchical RRCF model using a sliding window, and to generate long-term degradation detection results within the aggregation period. The distribution modeling module is used to input short-term anomaly detection results and long-term degradation detection results into the risk distribution model, calculate statistics and establish a distribution function, determine dynamic thresholds at quantiles and form a dynamic threshold set. The structure update module is used to update the tree structure when inserting or deleting high-dimensional time-series vectors, adjust the number and depth of trees when the rate of change of speed signal and load signal exceeds the threshold, perform subtree reconstruction when the tree structure is unbalanced, and recalculate the deviation score and update the risk distribution model and dynamic threshold set after the update. The consensus aggregation module is used to run a hierarchical RRCF model in each intelligent entity to calculate the local risk parameter set. The consensus module generates a consistent risk index, which is compared with a dynamic threshold set. When the index exceeds the threshold, a gearbox warning message is output and a timestamp is recorded.

2. The gearbox real-time monitoring and intelligent early warning system based on multi-agent technology according to claim 1, characterized in that, The modules are connected in the following way: The gearbox vibration signal, acoustic emission signal and speed signal are collected, and time synchronization and normalization processing is performed to obtain a standardized monitoring sequence. Time delay embedding is performed on standardized monitoring sequences to form a high-dimensional time series vector set; An RRCF model is constructed on a high-dimensional time-series vector set. Topological relationship constraints of gears, drive shafts and bearings are introduced during the selection of cutting dimensions. Non-uniform cutting is performed based on quantile density, and the deviation fraction sequence is output. The RRCF model is set as a hierarchical structure. The first layer processes the deviation score sequence and outputs short-term anomaly detection results. The second layer aggregates the short-term anomaly detection results within the time window and outputs long-term degradation detection results. Input the short-term anomaly detection results and long-term degradation detection results into the distribution modeling module to establish a risk distribution model and calculate a dynamic threshold set; When inserting and deleting high-dimensional time-series vectors, the number and depth of trees are adjusted according to the rotation speed and load signals. When the structure is unbalanced, node updates and subtree reconstruction are performed, and the risk distribution model and dynamic threshold set are updated simultaneously. Each agent independently runs a hierarchical RRCF model and outputs local risk parameters. All agents aggregate local risk parameters based on a consensus algorithm to generate a consistent risk index, and output gearbox warning information based on the consistent risk index and a dynamic threshold set.

3. The gearbox real-time monitoring and intelligent early warning system based on multi-agent technology according to claim 2, characterized in that, The generation of the high-dimensional time-series vector set specifically includes: Set the embedding dimension to a positive integer, the delay time to a sample interval of not less than one, the window length to the product of the embedding dimension and the delay time, and the step size to a fixed number of samples. Sliding truncation is performed on the standardized monitoring sequence according to the window length and step size. At the beginning of each window, observations of the number of embedded dimensions are extracted sequentially at fixed intervals of delay time to form a delayed sampling sequence. Delay sampling is performed on vibration signal, acoustic emission signal and rotation speed signal respectively. The three types of delay sampling results are spliced ​​in the order of vibration signal, acoustic emission signal and rotation speed signal. Within each type of signal, they are arranged in the order from zero delay to maximum delay to generate a single high-dimensional time vector. The current window is discarded when the starting point of the window cannot cover the complete number of embedded dimensions. For each high-dimensional time series vector, record the start timestamp, end timestamp, channel order, and delay order, and establish the correspondence between vector index and timestamp; All high-dimensional time series vectors are combined in chronological order to form a high-dimensional time series vector set, which is used to construct the input of the RRCF model.

4. The gearbox real-time monitoring and intelligent early warning system based on multi-agent technology according to claim 2, characterized in that, The construction and processing of the RRCF model specifically includes: An RRCF model is constructed on a high-dimensional time series vector set, with the number of trees set to a fixed integer and the maximum depth of each tree set to a fixed integer. When selecting the cutting dimension, the topological relationship between gears, drive shafts and bearings is represented as an adjacency matrix. Based on the adjacency matrix, the feature dimensions with connection relationships are divided into candidate sets, and a dimension is randomly selected from the candidate set as the cutting dimension. When determining the cut point, the range of values ​​of the selected cut dimension is used as the cut object. The range of values ​​is divided according to the quantiles. The interval density is calculated based on the number of observations in each quantile interval. The interval density is converted into sampling weights. The interval is selected according to the sampling weights. The cut point is randomly determined within the interval. After determining the cutting dimension and cutting point, the space where the high-dimensional time series vector is located is divided into two subspaces. Based on the relationship between the value of the high-dimensional time series vector in the cutting dimension and the size of the cutting point, the high-dimensional time series vector is assigned to the corresponding subspace. During the recursive partitioning process, the selection of cutting dimensions and determination of cutting points are repeatedly executed until the maximum depth of the tree or the capacity limit of the leaf nodes is reached. When inserting vectors into the tree structure, record the cutting path length and cutting depth, and determine the deviation score based on the difference between the cutting path length and the expected path length calculated based on the number of leaf node samples. After the insertion operation is completed in all trees, the deviation scores of the same high-dimensional time series vector in all trees are averaged to generate a deviation score sequence.

5. A real-time monitoring and intelligent early warning system for gearboxes based on multi-agent technology according to claim 2, characterized in that, The hierarchical processing of the RRCF model specifically includes: In the hierarchical RRCF model, define the first and second layers; In the first layer, the sliding window length is set to a fixed integer. The deviation score sequence is divided according to the sliding window order. The arithmetic mean of all deviation scores is calculated in each sliding window. The variance of all deviation scores is calculated. The maximum value of all deviation scores is calculated. The arithmetic mean, variance and maximum value are combined to form the short-term anomaly detection result. In the second layer, the short-term anomaly detection results are arranged in chronological order. The aggregation period is set to a fixed integer. The short-term anomaly detection results within one aggregation period are used as input. Within this input, the average of all arithmetic means, the average of all variances, and the maximum of all maximum values ​​are calculated. The average and maximum results are combined to form the long-term degradation detection results.

6. The gearbox real-time monitoring and intelligent early warning system based on multi-agent technology according to claim 2, characterized in that, The establishment of the risk distribution model and calculation of the dynamic threshold set specifically includes: Input the short-term anomaly detection results and long-term degradation detection results into the distribution modeling module; In the distributed modeling module, short-term anomaly detection results and long-term degradation detection results are collected within a fixed time range to form a risk observation sample set; Calculate the mean, variance, and quantiles of the risk observation sample set, and use the mean and variance as the initial statistics for distribution modeling; When establishing a risk distribution model, the relationship between the probability density function, cumulative distribution function, and dynamic threshold is defined as follows: ； in, Indicates the risk observation value. Represents a set of parameters. The symbol represents the energy function, and the integral symbol indicates a normalized calculation over the entire space. Represents the probability density function. Represents the cumulative distribution function. Indicates quantiles, This represents the dynamic threshold corresponding to the quantile. It represents the inverse function of the cumulative distribution function; All dynamic thresholds are combined into a dynamic threshold set, a timestamp is recorded for each dynamic threshold, and a correspondence is established with the risk observation sample set.

7. The gearbox real-time monitoring and intelligent early warning system based on multi-agent technology according to claim 2, characterized in that, The updating of the risk distribution model and the dynamic threshold set specifically includes: When a high-dimensional time series vector is inserted into the set, the newly added high-dimensional time series vector is sequentially input into all tree structures. During the input process, the path nodes are determined according to the cutting dimension and cutting point, and the timestamp of the newly added high-dimensional time series vector is registered at the leaf node position. When a deletion occurs in the high-dimensional time-series vector set, the timestamp of the deleted high-dimensional time-series vector is used to locate the leaf node in the tree structure, the current leaf node record is removed, and the associated cutting path is updated. When the speed signal and load signal change, calculate the rate of change and compare it with a preset threshold. When the rate of change is greater than the threshold, increase the number of trees or increase the depth of the trees. When the rate of change is less than the threshold, decrease the number of trees or decrease the depth of the trees. After the number of trees or the depth of the trees are adjusted, the balance of the tree structure is checked. When the tree structure is unbalanced, subtree reconstruction is performed on the affected branches, the nodes are re-divided into subspaces, and the cutting dimensions and cutting points are reallocated to the subspaces. After completing the insertion, deletion and tree structure adjustment of high-dimensional time series vectors, the deviation score is recalculated and the new deviation score is input into the risk distribution model; After the risk distribution model is updated, the dynamic threshold set is recalculated, a timestamp is recorded for each dynamic threshold, and a correspondence between the dynamic threshold and the risk observation sample set is established.

8. A real-time monitoring and intelligent early warning system for gearboxes based on multi-agent technology according to claim 2, characterized in that, The generation of the early warning information specifically includes: Run a hierarchical RRCF model in each intelligent body, input the local bias score sequence, and obtain the local short-term anomaly detection results, local long-term degradation detection results, and local risk distribution model; In each intelligent entity, a local risk parameter set is calculated based on the local risk distribution model, and the local risk parameter set is transmitted to the consensus module; In the consensus module, the local risk parameter set transmitted by all intelligent agents is received, consensus calculation is performed, and the consensus risk index is output. The consistency risk index is compared with a dynamic threshold set. When the consistency risk index is greater than the threshold in the dynamic threshold set, a gearbox warning message is generated. When generating gearbox warning information, record the consistency risk index value, the triggered dynamic threshold and the corresponding timestamp, and output the warning information.

Images