A key component thermal anomaly detection method based on improved isolation forest algorithm

Abstract

The present invention belongs to the field of machine learning fault prediction, and in particular relates to a method for detecting thermal anomalies of key components based on an improved isolation forest algorithm. First, temperature sensors and thermal imagers are deployed on the surfaces of key elevator components (such as traction machines, braking systems, etc.) to collect data. Then, an improved dynamic convolution kernel is combined with reinforcement learning to achieve data fusion. The convolution kernel parameters are determined by constructing an environment and designing a reward function, and then the features are integrated using an attention-based weighted fusion method. Then, the improved isolation forest algorithm is used for modeling and training, and the adaptive segmentation strategy and multi-granularity path length aggregation evaluation are improved. After modeling, heat dissipation modeling and dynamic threshold calibration are performed, and adversarial samples with heat dissipation noise are injected. Finally, the trained model is used to detect anomalies. This method can make full use of multi-source data, accurately extract features, adapt to changes in the operating state of the elevator, improve the accuracy and timeliness of thermal anomaly detection, and ensure the safe operation of the elevator.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning fault prediction, and in particular relates to a key component thermal anomaly detection method based on an improved isolation forest algorithm. Background Art

[0002] Elevators are important vertical transportation tools, but their key components are prone to overheating during long-term operation, causing safety accidents and impacting passenger safety and elevator efficiency. Traditional detection methods have significant shortcomings: data fusion methods are simple, often using direct splicing or fixed convolution kernel fusion. These methods fail to account for dynamic data changes, making it difficult to mine potential data relationships. The fused features are weak in characterizing thermal anomalies, resulting in poor detection accuracy and timeliness. When using the isolation forest algorithm for anomaly detection, the traditional construction method randomly selects features and segmentation points, failing to consider data distribution characteristics and feature correlations. It relies solely on simple path length assessments, which can easily miss potential anomalies and result in low detection accuracy and reliability. Furthermore, heat dissipation in elevator shafts is affected by multiple factors. Traditional models use fixed thresholds, which cannot adapt to heat fluctuations, are prone to misjudgments, and cannot effectively distinguish between true faults and environmental heat dissipation issues. Summary of the Invention

[0003] In response to the technical problems existing in the above background technology, the present invention proposes a key component thermal anomaly detection method based on an improved isolation forest algorithm.

[0004] In order to achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0005] S1. First, deploy temperature sensors and thermal imagers on the surfaces of key elevator components.

[0006] S2. Use improved dynamic convolution kernel to achieve data fusion

[0007] S3. Use the improved isolation forest algorithm to build a model and use the original data to train the model. The improvements of the isolation forest are to improve the adaptive segmentation strategy and evaluate the multi-granularity path length aggregation;

[0008] S4. Finally, use the trained model to detect anomalies in the temperature data of key elevator components.

[0009] The specific implementation method of step S2 using the improved dynamic convolution kernel to achieve data fusion is:

[0010] S21, sort the collected temperature data in chronological order T=[T1, T2, ...T n ], the thermal image collected by the thermal imager is recorded as I=[I1,I2,...I m ], where n is the length of the time series, m is the number of frames of the thermal image, and the thermal image is normalized;

[0011] S22, then build a reinforcement learning environment for the dynamic convolution kernel, and use the currently processed temperature data fragment and thermal image frame as state S t , action a t Represents the parameters of the dynamic convolution kernel;

[0012] S23. Design the reward function. The reward function is defined as:

[0013] Among them, α1, α2, α3, α4 are weight coefficients, FD(F), CP(a t ), CA(V t ,L t ,F) are respectively the accuracy, feature diversity metric, the complexity of the convolution kernel corresponding to the action, and the operating status of the key components of the elevator;

[0014] S24, then conduct reinforcement learning training to learn to select the optimal dynamic convolution kernel parameters under different states;

[0015] S25. At each time step t, determine the convolution kernel parameters for temperature data and thermal image data feature extraction based on the learned action;

[0016] S26. Finally, the extracted data features are fused.

[0017] Preferably, the key parts in step S1 include a traction machine, a braking system, a control cabinet, cables and guide rails.

[0018] Preferably, the dynamic convolution kernel and parameters in step S22 include the size, step size and padding of the convolution kernel.

[0019] As a preference, in step S23, CA(V t ,L t ,F) represents the effect of feature F on the current running speed V t and load condition L t The fitness of is calculated as follows: Among them, Correlation(F,V t ,L t ) is a measure of the characteristic F and the running speed V t and load L t The correlation was obtained by Pearson's correlation coefficient.

[0020] As a preference, the specific fusion method in step S26 is a weighted fusion method based on attention, and the attention weights ω of the thermal image data features and the temperature data are calculated respectively by the Softmax function. T ,ωI , and then perform weighted fusion to obtain the final feature in Stands for element-wise multiplication.

[0021] Preferably, the improved isolation forest in step S3 is implemented as follows:

[0022] S31, first initialize the isolation forest model;

[0023] S32. Then, for each feature dimension in the sample subset, kernel density estimation is used to calculate the data density: Where h is the bandwidth and H is the Epanechnikov kernel function;

[0024] S33. On the selected feature dimension, find the segmentation point s that maximizes the density difference of the subsets after segmentation. * ,Right now, where s - , s + Respectively represent the left and right areas of the split point;

[0025] S34, repeating the splitting step for the left and right child nodes obtained by splitting until the depth of the tree reaches a preset maximum depth;

[0026] S35. Construct multiple isolated trees according to the set number of trees to form a complete isolation forest model;

[0027] S36. Finally, improve the anomaly score formula: Where L(x) is the path length, c(n) is the normalization factor, is the depth weighting function, where d0 is the preset depth.

[0028] Preferably, after modeling using the improved isolation forest algorithm and before anomaly detection, there are steps of heat dissipation modeling and dynamic threshold calibration:

[0029] Construct an elevator shaft heat dissipation model and define the heat dissipation coefficient h vent for: Among them A open is the area of ​​the shaft ventilation opening, V shaft is the well volume, ΔP airflow is the natural convection intensity driven by pressure difference, k1 and k2 are experimental calibration coefficients;

[0030] And introduce the time integral term to correct the temperature threshold T thres , the calculation formula is: Among them, P heat (τ) is the heat generation power of the component.

[0031] As a preferred approach, an adversarial sample with heat dissipation noise is injected into the improved isolation forest model. The calculation formula is: Where T real is the true temperature, ε is the disturbance intensity, is the gradient sign of the model loss function with respect to the temperature data T.

[0032] Compared with the prior art, the advantages and positive effects of the present invention are: DETAILED DESCRIPTION

[0033] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described below in conjunction with the embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0034] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0035] In today's high-rise cities, elevators, as a vital means of vertical transportation, are of paramount importance for their safety. Key elevator components, such as traction motors, braking systems, control cabinets, cables, and guide rails, are highly susceptible to safety accidents caused by overheating during long-term operation. According to incomplete statistics, failures caused by thermal anomalies in elevator components account for a significant proportion of total failures, seriously impacting passenger safety and the normal operation efficiency of the elevator. To effectively address the shortcomings of traditional elevator thermal anomaly detection methods, this paper proposes a key component thermal anomaly detection method based on an improved isolation forest algorithm.

[0036] First, high-precision temperature sensors and high-resolution thermal imagers were deployed on the elevator's traction motor, braking system, control cabinet, cables, and guide rails. These devices were strategically located in heat-prone areas of key components to ensure accurate temperature and thermal image data collection.

[0037] Considering the relative simplicity of traditional data fusion methods, temperature data and thermal image data are typically directly concatenated, or feature extraction is performed using a fixed convolution kernel before fusion. The parameters of these fixed convolution kernels are pre-set and do not adjust based on dynamic data changes. This connection method fails to consider the dynamic nature of the data, resulting in incomplete feature extraction. Furthermore, simple data concatenation or fixed convolution kernel fusion methods fail to fully exploit the underlying relationships between the data and effectively integrate the complementary information between temperature and thermal image data. This results in insufficient characterization of thermal anomalies by the fused data features, thereby impacting the accuracy and timeliness of thermal anomaly detection. This invention, however, employs an improved dynamic convolution kernel combined with reinforcement learning. A reinforcement learning environment is constructed, using temperature data segments and thermal image frames as states and the parameters of the dynamic convolution kernel as actions. By designing a reward function that incorporates factors such as accuracy, feature diversity metrics, convolution kernel complexity, and the operating status of key elevator components, the algorithm is guided to learn and select the optimal dynamic convolution kernel parameters under different states. The collected temperature data is sorted in chronological order: T = [T1, T2, ... T n ], the thermal image collected by the thermal imager is recorded as I=[I1,I2,...I m ], where n is the length of the time series and m is the number of frames of the thermal image. The thermal image is normalized. Then, the reinforcement learning environment of the dynamic convolution kernel is constructed, and the currently processed temperature data fragment and thermal image frame are used as the state S t , action a t Represents the parameters of the dynamic convolution kernel, including the size, step size and padding of the convolution kernel. Next, the reward function is designed, which is defined as: Among them, α1, α2, α3, α4 are weight coefficients, FD(F), CP(a t ), CA(V t ,L t ,F) are respectively the accuracy, feature diversity metric, the complexity of the convolution kernel corresponding to the action and the operating status of the key components of the elevator, where feature F is the current running speed V t and load condition L t The fitness of is calculated as follows: Among them, Correlation(F,V t ,L t ) is a measure of the characteristic F and the running speed V t and load L tThe correlation is obtained by the Pearson correlation coefficient. Then reinforcement learning training is performed to learn to select the optimal dynamic convolution kernel parameters under different states. At each time step t, the convolution kernel parameters are determined based on the learned action for temperature data and thermal image data feature extraction. Finally, the extracted data features are weighted fused based on attention, and the attention weights ω of the thermal image data features and temperature data are calculated respectively through the Softmax function. T ,ω I , and then perform weighted fusion to obtain the final feature in "" stands for element-by-element multiplication. This improvement allows the system to adapt to changes in the elevator's operating state, dynamically adjusting convolution kernel parameters to accurately extract temperature and thermal image data features under different operating conditions. For example, it can find the most appropriate convolution kernel parameters for both lightly and heavily loaded elevators, comprehensively capturing changes in component thermal states. Furthermore, the dynamic convolution kernel, trained using reinforcement learning, can deeply explore potential relationships between data, improving the ability of fused data features to characterize thermal anomalies and enhancing the accuracy and timeliness of thermal anomaly detection.

[0038] Then, for traditional anomaly detection in similar scenarios, if the isolation forest algorithm is used, its implementation principle is to construct an isolation tree based on randomly selected features and split points. During the construction process, samples are randomly selected from the data set to build a tree. For each node, a feature dimension is randomly selected, and then a split point is randomly selected within the value range of the dimension to divide the data set into two left and right subsets. This process continues until the preset tree depth is reached. Finally, the degree of abnormality of the sample is evaluated by calculating the path length of the sample in the tree. The shorter the path length, the more likely the sample is an anomaly. However, this traditional method does not take into account the distribution characteristics of the data and the association between different features. On the other hand, only evaluating anomalies based on simple path length does not make full use of multi-granularity information, and it is easy to miss some potential abnormal situations, which greatly reduces the accuracy and reliability of anomaly detection. The improvement of the present invention is in the adaptive segmentation strategy. The data density is calculated by kernel density estimation, and the split point that maximizes the density difference of the subsets after segmentation is found on the selected feature dimension. Compared with the traditional random selection of split points, this method can better segment according to the distribution characteristics of the data itself and accurately grasp the key features. In terms of multi-granularity path length aggregation evaluation, we improve the anomaly score formula and introduce a deep weighting function. This comprehensively considers multiple factors such as path length and normalization factors, evaluates data at multiple granularities, and comprehensively captures abnormal patterns in the data. First, the isolation forest model is initialized; then, for each feature dimension in the sample subset, kernel density estimation is used to calculate the data density: Where h is the bandwidth and H is the Epanechnikov kernel function; on the selected feature dimension, find the segmentation point s that maximizes the density difference of the subsets after segmentation * ,Right now, where s - , s + Represent the left and right areas of the segmentation point respectively; repeat the segmentation steps for the left and right child nodes obtained by segmentation until the depth of the tree reaches the preset maximum depth; construct multiple isolated trees according to the set number of trees to form a complete isolation forest model; finally, improve the anomaly score formula: Where L(x) is the path length, c(n) is the normalization factor, is a depth-weighted function, where d0 is the preset depth. This improvement enhances the model's stability and generalization capabilities, reducing false positives and missed detections. Data fusion provides the isolation forest algorithm with optimized, highly representative feature data for thermal anomalies, making algorithm training more targeted and improving its ability to identify abnormal data. The isolation forest algorithm then effectively analyzes the fused data to accurately identify thermal anomalies.

[0039] In addition, considering that the elevator shaft is a semi-enclosed space, the heat dissipation efficiency is easily affected by factors such as ambient temperature and vent blockage. The traditional model is prone to misjudgment due to heat dissipation fluctuations under a fixed threshold. Therefore, after modeling, the heat dissipation modeling and dynamic threshold calibration steps are also performed: the elevator shaft heat dissipation model is constructed, and the heat dissipation coefficient h is defined. vent for: Among them A open is the area of ​​the shaft ventilation opening, V shaft is the well volume, ΔP airflow is the natural convection intensity driven by the pressure difference, k1 and k2 are experimental calibration coefficients; and the time integral term is introduced to correct the temperature threshold T thres , the calculation formula is: Among them, P heat (τ) is the heat generation power of the component. The adversarial sample with heat dissipation noise injected into the improved isolation forest model is calculated as follows: Where T real is the true temperature, ε is the disturbance intensity, is the gradient sign of the model loss function with respect to the temperature data T. By simulating heat dissipation anomalies (such as high temperatures in summer and dust accumulation in vents) through adversarial samples, the model learns to distinguish between real component failures and environmental heat dissipation failures during the training phase. For example: when the ventilation efficiency of the shaft decreases, the adversarial sample generates simulated temperature rise data, forcing the model to dynamically adjust the threshold to avoid false alarms of temperature rises of normal components. In an embodiment, the model accurately distinguishes between overheating of cable joints (real anomalies) and ambient temperature rise (non-abnormalities) under high temperatures in summer through adversarial training. In this step, the time integral term is introduced in combination with the component heat generation power to correct the temperature threshold, which can adapt to heat generation changes in real time and avoid misjudgments due to heat dissipation fluctuations. Heat dissipation noise adversarial samples are injected into the model to simulate heat dissipation anomalies, so that the model can learn to distinguish between real faults and environmental heat dissipation problems during training. For example, real anomalies such as overheating of cable joints can be accurately judged under high temperatures in summer.

[0040] Finally, the trained model is used to detect anomalies in the temperature data of key elevator components, providing more reliable anomaly detection guarantees for the safe operation of the elevator and enhancing the efficiency of the entire thermal anomaly detection.

[0041] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any person skilled in the art may utilize the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes for application in other fields. However, any simple modification, equivalent change, and modification of the above embodiments made in accordance with the technical essence of the present invention without departing from the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A key component thermal anomaly detection method based on an improved isolation forest algorithm, characterized in that: The following steps are involved: S1. First, deploy temperature sensors and thermal imagers on the surfaces of key elevator components. S2, using improved dynamic convolution kernel to achieve data fusion; S3. Use the improved isolation forest algorithm to build a model and use the original data to train the model. The improvements of the isolation forest are to improve the adaptive segmentation strategy and evaluate the multi-granularity path length aggregation; S4. Finally, use the trained model to detect anomalies in the temperature data of key elevator components. The specific implementation method of step S2 using the improved dynamic convolution kernel to achieve data fusion is: S21, sort the collected temperature data in chronological order T=[T1, T2, ...T n ], the thermal image collected by the thermal imager is recorded as I=[I1,I2,...I m ], where n is the length of the time series, m is the number of frames of the thermal image, and the thermal image is normalized; S22, then build a reinforcement learning environment for the dynamic convolution kernel, and use the currently processed temperature data fragment and thermal image frame as state S t , action a t Represents the parameters of the dynamic convolution kernel; S23. Design the reward function. The reward function is defined as: Among them, α1, α2, α3, α4 are weight coefficients, CP(a t ), CA(V t ,L t ,F) are respectively the accuracy, feature diversity metric, the complexity of the convolution kernel corresponding to the action, and the operating status of the key components of the elevator; S24, then conduct reinforcement learning training to learn to select the optimal dynamic convolution kernel parameters under different states; S25. At each time step t, determine the convolution kernel parameters for temperature data and thermal image data feature extraction based on the learned action; S26. Finally, the extracted data features are fused.

2. The key component thermal anomaly detection method based on the improved isolation forest algorithm according to claim 1 is characterized in that: The key parts in step S1 include the traction machine, the braking system, the control cabinet, the cables and the guide rails.

3. The key component thermal anomaly detection method based on the improved isolation forest algorithm according to claim 1 is characterized in that: The dynamic convolution kernel and parameters in step S22 include the size, step size and padding of the convolution kernel.

4. The key component thermal anomaly detection method based on the improved isolation forest algorithm according to claim 1 is characterized in that: In step S23, CA(V t ,L t ,F) represents the effect of feature F on the current running speed V t and load condition L t The fitness of is calculated as follows: Among them, Correlation(F,V t ,L t ) is a measure of the characteristic F and the running speed V t and load L t The correlation was obtained by Pearson's correlation coefficient.

5. The key component thermal anomaly detection method based on the improved isolation forest algorithm according to claim 1 is characterized in that: The specific fusion method in step S26 is a weighted fusion method based on attention, and the attention weights ω of the thermal image data features and temperature data are calculated respectively by the Softmax function. T ,ω I , and then perform weighted fusion to obtain the final feature in Stands for element-wise multiplication.

6. The key component thermal anomaly detection method based on the improved isolation forest algorithm according to claim 1 is characterized in that: The implementation of the improved isolation forest in step S3 is: S31, first initialize the isolation forest model; S32. Then, for each feature dimension in the sample subset, kernel density estimation is used to calculate the data density: Where h is the bandwidth and H is the Epanechnikov kernel function; S33. On the selected feature dimension, find the segmentation point s that maximizes the density difference of the subsets after segmentation. * ,Right now, where s - , s + Respectively represent the left and right areas of the split point; S34, repeating the splitting step for the left and right child nodes obtained by splitting until the depth of the tree reaches a preset maximum depth; S35. Construct multiple isolated trees according to the set number of trees to form a complete isolation forest model; S36. Finally, improve the anomaly score formula: Where L(x) is the path length, c(n) is the normalization factor, is the depth weighting function, where d0 is the preset depth.

7. According to the method for detecting thermal anomalies of key components based on the improved isolation forest algorithm in claim 1, after modeling using the improved isolation forest algorithm and before anomaly detection, there are further steps of heat dissipation modeling and dynamic threshold calibration: Construct an elevator shaft heat dissipation model and define the heat dissipation coefficient h vent for: Among them A open is the area of ​​the shaft ventilation opening, V shaft is the well volume, ΔP airflow is the natural convection intensity driven by pressure difference, k1 and k2 are experimental calibration coefficients; And introduce the time integral term to correct the temperature threshold T thres , the calculation formula is: Among them, P heat (τ) is the heat generation power of the component.

8. According to the method for detecting thermal anomalies of key components based on the improved isolation forest algorithm of claim 7, an adversarial sample with heat dissipation noise is injected into the improved isolation forest model. The calculation formula is: Where T real is the true temperature, ε is the disturbance intensity, is the gradient sign of the model loss function with respect to the temperature data T.

Images