Equipment status perception system based on peak-shifting and dangerous situation perception

Abstract

The present invention discloses an equipment state perception system based on peak shifting and dangerous situation perception, which relates to the field of electrical digital data processing technology, including a data acquisition unit, a model training and optimization unit, a situation perception unit, a peak shifting control unit, a system interface and a user interaction unit. The present invention uses a Gaussian kernel function to dynamically calculate the similarity of equipment state data features, adjusts the similarity matrix in combination with a gradient update term, dynamically assigns weights in feature extraction, and uses adaptive discard probability to reduce redundant features, thereby improving the feature extraction efficiency of the model in a high-dimensional feature space and solving the problem that too many redundant features of equipment state data affect the dimensionality reduction effect. The equipment state features are extracted layer by layer using a deep forest algorithm with a multi-layer random forest structure, and the decision tree is optimized in combination with a dynamic weight adjustment mechanism and an enhanced learning strategy, thereby enhancing the model's ability to classify equipment state changes and abnormal states.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical digital data processing, and in particular to an equipment state perception system based on peak shifting and dangerous situation perception. Background Art

[0002] With the increasing complexity and automation level of industrial equipment, the monitoring and management of equipment status have become increasingly important in industrial production. The operating status of equipment directly affects production efficiency, energy consumption and operating costs. Traditional equipment status monitoring methods mostly rely on fixed rules or empirical models, which are difficult to adapt to the diverse and dynamic operating environment. In addition, the operating data sources of industrial equipment are complex, including real-time monitoring data, historical record data, and environmental sensor data. The timeliness, quality and diversity of these data put forward higher requirements for equipment status perception. However, current technology is difficult to effectively integrate multi-source data, and often due to problems such as data redundancy and noise interference, the status prediction is not accurate enough.

[0003] In practical applications, the complexity of equipment status is also reflected in the expression of high-dimensional data features. Continuous time series signals such as vibration, current, and voltage often contain rich status information, but traditional dimensionality reduction and analysis methods are difficult to balance feature extraction efficiency and information integrity. In addition, equipment status may change dynamically during operation. Traditional classification models are insufficient in capturing dynamic features and have limited detection effects on potential faults and abnormal patterns. At the same time, equipment maintenance strategies are often based on post-event response, lacking predictive intervention measures for potential faults and risks, which can easily lead to unplanned downtime and high maintenance costs.

[0004] In the prior art, a Chinese invention patent with publication number CN119180633A proposes a method and system for abnormal warning of equipment based on machine learning, including collecting working data of equipment, preprocessing the working data, screening the operating data to obtain abnormal data, importing the abnormal data into an abnormal function to obtain an abnormal coefficient, decomposing the abnormal coefficient and the state data characteristics to obtain a feature vector, comprehensively evaluating the state data to obtain a state evaluation value, calculating the similarity between the abnormal coefficient and the state evaluation value to obtain an abnormal state degree, constructing an abnormal warning model for equipment based on machine learning according to the feature vector and the abnormal state degree, inputting the data to be evaluated into the abnormal warning model for equipment to obtain an abnormal warning value for equipment, and performing an abnormal warning according to the abnormal warning value for equipment. A Chinese invention patent with publication number CN119150083A proposes a method and system for analyzing the operating state of industrial equipment based on machine learning, which relates to the field of equipment detection technology. Acquire multiple types of detection items and abnormal behaviors, combine them respectively, and input the corresponding data in the historical detection records and historical abnormal records into different deep networks for training, and select several detection items that are sensitive to each abnormal behavior; obtain the type of abnormal behavior of the target to be monitored; obtain the parameters of the detection items to be monitored; according to the type of abnormal behavior of the target to be monitored and the several detection items that are sensitive to each abnormal behavior, identify and analyze the parameters of the detection items to be monitored to obtain the state parameters of the abnormal behavior of the target to be monitored. The Chinese invention patent with publication number CN119148535A proposes a method and system for intelligent device control based on machine learning, which involves the field of intelligent device control technology, including S1, data acquisition; S2, model training; S3, prediction and control; S4, interactive control; S5, feedback regulation. This intelligent device control method and system based on machine learning uses machine learning technology to analyze and learn user behavior and device usage data, thereby predicting user behavior habits and realizing intelligent control of intelligent devices. It not only simplifies user operation steps and reduces user difficulty of use, but also can automatically adapt to user behavior habits, adjust control strategies according to user behavior changes, optimize device operation modes and parameters, and keep pace with user needs.

[0005] The existing technology has the following shortcomings: 1. In the task of equipment status perception, the data collection and processing methods are limited, and it is impossible to effectively integrate real-time monitoring and historical record data, resulting in difficulty in ensuring the timeliness and integrity of equipment status data; 2. In the task of equipment status perception, the traditional dimensionality reduction method has insufficient suppression ability for redundant features and high-dimensional noise of equipment status data, and cannot take into account both feature extraction efficiency and retention of important information; 3. In the task of equipment status perception, conventional classification algorithms have poor adaptability to dynamically changing equipment status data, and lack the ability to effectively identify multi-level features and abnormal patterns, resulting in unstable classification performance; 4. In the task of equipment status perception, the equipment predictive maintenance strategy only relies on simple rules or fixed models, and fails to formulate graded response measures in combination with the dynamic equipment status perception results, and the risks of peak overload and potential failures are not effectively avoided. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide an equipment status perception system based on peak-shifting and dangerous situation perception.

[0007] The technical solution adopted to solve the above technical problems is: an equipment status perception system based on peak-shifting and dangerous situation perception, including: a data acquisition unit, which is used to collect multi-source data and manually annotate the collected data. The annotated categories include: normal: indicating that the equipment is in normal operation; warning: indicating that there is a potential risk in the operation of the equipment; fault: indicating that there is a fault in the operation of the equipment; a model training and optimization unit, which processes the equipment status training data through the equipment status data feature dimensionality reduction model, and adopts the autoencoder algorithm based on adaptive discarding as the equipment status data feature dimensionality reduction model; the reduced features are input into the equipment status perception model for equipment status recognition, and a deep forest algorithm based on dynamic learning is adopted as The equipment status perception model improves the accuracy of classification, extracts features layer by layer through the hierarchical structure of cascaded random forests, and gradually improves the expression ability of the model; the situation awareness unit applies the trained model to actual data input to perform classification, anomaly detection or fault prediction of equipment status perception; the peak-shifting control unit optimizes the equipment's operating time and workload through peak-shifting strategies based on the timing characteristics of equipment load and the prediction results of operating status, thereby reducing the risk of equipment overload and failure during peak hours; the system interface and user interaction unit provides a user-friendly interactive interface, supports status visualization, alarm notification and policy adjustment, and through a graphical interface, users can intuitively view equipment operating status, dangerous situation warnings and peak-shifting control suggestions.

[0008] Furthermore, the data attributes of the collected data include: equipment operating time Ra, equipment load percentage Da, ambient temperature Pa, ambient humidity Qa, equipment current Sa, equipment voltage Ta, equipment vibration frequency Ua, equipment noise level Va, preventive maintenance times Wa, and failure times Xa.

[0009] Furthermore, the training process of the self-encoder algorithm based on adaptive discarding includes the following steps:

[0010] S101, in the initial stage of dimensionality reduction of device state data features, firstly, by calculating the similarity between the input device state data feature vectors, the relationship between each device state data feature is dynamically evaluated, and the correlation relationship between the device state data features is established by preliminarily calculating and updating the feature similarity of the input device state data, so as to focus on the main device state data features with precise weights in the subsequent steps, and suppress irrelevant device state data features. The device state data feature similarity is initially measured by a Gaussian kernel function, and a gradient adjustment term is used to realize a dynamic update mechanism. The calculation method is expressed as:

[0011]

[0012] In the formula, is the adjustment item in the similarity matrix, representing the The samples and The change of samples at the feature gradient level, To control the parameters of the gradient update weights, and Respectively and The characteristic gradient of each sample is used to characterize the changing trend of the local characteristics of the device status data. is the first sample index input into the autoencoder, is the sample index of the second input into the autoencoder;

[0013] Based on the above adjustment items, the similarity matrix is ​​updated and the calculation method is expressed as:

[0014]

[0015] In the formula, For the The samples and The similarity of the updated samples is is the similarity calculation function, For the The samples and The square of the Euclidean distance between samples is used to characterize the difference between the two in the feature space of the original device status data. is the L2 norm, is the first input to the autoencoder samples, is the first input to the autoencoder samples, The smoothing parameter is the Gaussian kernel control coefficient;

[0016] S102, training the adaptive encoding and decoding network of the autoencoder, mapping the high-dimensional device status data features to a low-dimensional space and reconstructing them through the autoencoder structure, using a weighted reconstruction error function based on similarity to enhance the network's attention to key device status data features, retaining important information while reducing the dimensionality. During the training process of the autoencoder, the calculation method of the weighted reconstruction loss function is expressed as:

[0017]

[0018] In the formula, is the weighted reconstruction error of the autoencoder, is the number of samples input to the autoencoder for the current batch, For the The discard status of samples, For the The weighting coefficient of samples is used to strengthen the focus on important features. is the first input to the autoencoder samples, representing the The original device status data feature vector of samples, For the The device status data feature vector reconstructed by samples is For distance loss;

[0019] In order to balance the similarity and the distance between samples, the weighted coefficient is calculated based on the similarity matrix, and the calculation method is expressed as:

[0020]

[0021] In the formula, To determine the trade-off coefficient between similarity and distance;

[0022] S103, through nonlinear mapping and re-weighting operations, the characteristic pattern of complex device status data is captured and amplified, and a nonlinear transformation is performed on the input device status data characteristic matrix, so as to better characterize the nonlinear structure of the original device status data in the new device status data characteristic space. The calculation method of nonlinear mapping is expressed as:

[0023]

[0024] In the formula, For the The nonlinear characteristics after sample mapping, is the hyperbolic tangent function, which is used to simulate nonlinear mapping. is the weight matrix of the autoencoder, For the The noise intensity of each sample, is the bias term of the autoencoder;

[0025] In order to enhance the contribution of nonlinear features to reconstruction and dimensionality reduction, the distance loss is calculated as follows:

[0026]

[0027] In the formula, For the The feature weighting coefficient of samples in the nonlinear feature space, is a parameter that controls the sensitivity of class distinction. For the The Euclidean distance between the nonlinear features of a sample and its true category, is the Euclidean distance calculation function;

[0028] The feature reweighting coefficient of equipment status data can amplify important features in nonlinear space, and the calculation method is expressed as:

[0029]

[0030] In the formula, is the similarity calculation function, For the The similarity measurement result between a sample and other samples is To control the parameters of the similarity influence range, For the The feature change gradient of a sample is used to reflect the importance of the feature in the update direction;

[0031] The non-stationary noise suppression method based on physical modeling is adopted to estimate the dynamic distribution of noise in the feature space of equipment status data through physical parameter modeling, and suppress the noise during dimensionality reduction. The calculation method is expressed as:

[0032]

[0033] In the formula, For the The noise intensity of each sample, is the noise baseline intensity, is the noise amplitude coefficient, which measures the range of noise fluctuations. The frequency coefficient of is the parameter of the noise change speed, For the The position index of each sample in the feature space;

[0034] S104, for redundant features in the training process, an adaptive discard probability is used to control the retention and discarding of device status data features. The discard operation is represented by a binary random variable, and the calculation method is expressed as:

[0035]

[0036] In the formula, For the The discard status of a sample, which indicates whether the feature is retained;

[0037] The discard probability is calculated based on the original reconstruction error and the additional gain term, and the calculation method is expressed as:

[0038]

[0039] In the formula, For the The probability of discarding a sample is is the adjustment parameter of the discard probability, For the The reconstruction error of samples is For the The gain term of samples;

[0040] The sample gain term is used to strengthen the focus on important features, and the calculation method is expressed as:

[0041]

[0042] In the formula, For the The gain term of samples, For the The weighting coefficient of samples, is the smoothing coefficient, is the L2 norm;

[0043] S105, adopting the discard feedback method to modify the discard strategy according to the reconstruction situation after the training is completed, and the calculation method is expressed as:

[0044]

[0045] In the formula, is the feedback loss after discarding, For the The reconstruction error of samples is is the feedback weight, For the The discard status of each sample determines the impact of the discard strategy on model training;

[0046] S106 adopts the method of adaptive learning rate adjustment and optimal dimension reduction selection. By adjusting the learning rate in real time according to the changes in the reconstruction error and the discarding mechanism, local convergence is avoided and model training is accelerated. The optimal dimension reduction is selected by using information gain and feature variance evaluation to ensure that the device status data after dimension reduction still retains the core information and has good discrimination. The learning rate update method is expressed as:

[0047]

[0048] In the formula, is the learning rate of the autoencoder at the current iteration, is the learning rate of the autoencoder in the previous iteration, Adjust the step size for the learning rate of the autoencoder, is the average reconstruction error of this round of training, is an additional dynamic coefficient used to fine-tune the learning rate in combination with the dropout probability. Set to 0.5, Tuning parameters for the learning rate;

[0049] A comprehensive measurement method based on information gain and feature variance is used to select the optimal dimension reduction dimension. The calculation method is expressed as:

[0050]

[0051] In the formula, is the optimal dimension reduction dimension, For the The information gain of the device status data feature is used to measure its identifiability. is the index of the device status data feature, is the characteristic dimension of the equipment status data before dimensionality reduction, For the The variance of the device status data feature is used to measure its discrimination in the sample distribution;

[0052] The calculation method of feature information complementarity is used to measure the degree of complementarity between the features of equipment status data. The calculation method is expressed as:

[0053]

[0054] In the formula, For the The first The information complementarity between the characteristics of device status data, For the Shannon entropy of the device status data features, For the Shannon entropy of the device status data features, For the The device status data characteristics and The joint entropy of the device status data features;

[0055] S107, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0056] Furthermore, the training process of the deep forest algorithm based on dynamic learning includes the following steps:

[0057] S201, input the device status data after feature dimension reduction into the deep forest algorithm based on dynamic learning, establish a multi-layer deep forest structure, and assign initial weights to each random forest in the deep forest to improve the model's deep processing capability for nonlinear data. The weight calculation method for each layer of random forest is expressed as:

[0058]

[0059] In the formula, For the The weights of the random forest, is the index of the first random forest number, is the number of statistical items, is the adjustment coefficient of error rate to weight, For the The error rate corresponding to each statistical item;

[0060] S202, dynamically adjust the partial derivative of the random forest weight based on the loss function, so that the model can adapt to changes in data distribution more quickly. The deep forest model dynamically adjusts the partial derivative of the random forest weight so that it can quickly respond to changes in the distribution of device status data, which is expressed as:

[0061]

[0062] In the formula, For the The amount by which the random forest weights are adjusted, is the index of the second random forest number, is the learning rate of the random forest weights, is the number of random forests in the deep forest, For the The balance coefficient of the random forest loss is preset manually. Based on The loss function calculated by a random forest is is the symbol of partial derivative;

[0063] S203, the decision tree structure is self-optimized by using the reinforcement learning strategy, so that the model can still maintain a high recognition ability when facing variable and complex data distribution. Deep Forest dynamically optimizes the decision tree structure through the reinforcement learning strategy, so that the model can still maintain high classification performance when facing novel or abnormal device states, which is expressed as:

[0064]

[0065] In the formula, is the performance function parameter before updating, is the updated performance function parameter, is the learning step size of the performance function, is the gradient of the performance function, is the performance function, is the performance function parameter;

[0066] The performance function is based on an independent evaluation of each decision tree in the random forest and then a comprehensive score, and the calculation method is expressed as:

[0067]

[0068] In the formula, is the index of the number of decision trees, is the total number of decision trees in the random forest, For the The weight of a decision tree is preset by the user. For the The classification accuracy of a decision tree under the current parameters;

[0069] S204, a comprehensive method based on information gain and distribution characteristics is used to evaluate and screen features, reduce noise interference and improve the discriminative power of the overall model, which is expressed as:

[0070]

[0071] In the formula, For the Features, is the information gain function, For the dataset The entropy of is the dataset input to the deep forest, Features Value The probability of is the possible value of the feature, Values ​​is the feature value set, is the information gain regularization coefficient, Features In the dataset The variance on ;

[0072] According to the specific definition of variance, the information gain is corrected. The calculation method of variance is expressed as:

[0073]

[0074] In the formula, For the data points in the feature The value of is the index of the data point input to the deep forest, Features The mean of is the size of the dataset input to the deep forest;

[0075] S205, after completing the feature importance evaluation and selection, an adaptive feature transformation layer is further used to re-encode the input features more flexibly during the training process to enhance the generalization ability of the model in complex environments, which is expressed as:

[0076]

[0077] In the formula, is the adaptive feature transformation layer function, is the input feature vector of the deep forest, is the weight matrix of the adaptive feature transformation layer, is the bias vector of the adaptive feature transformation layer, is the nonlinear activation function of the adaptive feature transformation layer, and the activation function of the adaptive feature transformation layer adopts the ReLU activation function;

[0078] The feature transformation layer parameters are updated based on the loss function of feature importance and classification effect. The loss function calculation method of the feature transformation layer is expressed as:

[0079]

[0080] In the formula, is the loss function of the feature transformation layer, The number of samples input to the feature transformation layer for the current batch, The index of the sample input to the feature transformation layer for the current batch, is the cross entropy loss, For the The true labels of samples, For the model The predicted output of samples is is the regularization coefficient of the feature transformation layer, For Regularization function, specifically L2 regularization;

[0081] Parameter update is performed according to the back propagation algorithm, and the calculation method is expressed as:

[0082]

[0083]

[0084] In the formula, is the weight matrix of the updated adaptive feature transformation layer, is the weight matrix of the adaptive feature transformation layer before updating, is the bias vector of the updated adaptive feature transformation layer, is the bias vector of the adaptive feature transformation layer before updating, is the learning rate of the feature transformation layer, is the gradient of the loss function of the feature transformation layer with respect to the weight matrix of the adaptive feature transformation layer, is the gradient of the loss function of the feature transformation layer with respect to the bias vector of the adaptive feature transformation layer;

[0085] S206, in the inter-layer feedback mechanism stage, by allowing the high-level output and feedback signals of the deep forest to act together on the input of the next layer, the coordination between the layers and the ability to capture complex patterns are strengthened, which is expressed as:

[0086]

[0087] In the formula, For the The input of a random forest is For the The output of a random forest, For the The regulation signal fed back by the random forest is The nonlinear mapping result of a random forest is is the activation function of random forest;

[0088] The specific inter-layer interaction is realized based on the network layer that fuses the output and feedback signals, providing reliable technical support for equipment maintenance and monitoring. The activation function calculation method of random forest is expressed as:

[0089]

[0090] In the formula, is the Sigmoid activation function, For the The bias of a random forest.

[0091] The beneficial effects of the present invention are as follows: (1) In the device state perception task, the present invention uses the Gaussian kernel function to dynamically calculate the feature similarity of the device state data, combines the gradient update term to adjust the similarity matrix, dynamically assigns weights in feature extraction, and uses adaptive discarding probability to reduce redundant features, thereby improving the feature extraction efficiency of the model in the high-dimensional feature space and solving the problem that too many redundant features of the device state data affect the dimensionality reduction effect.

[0092] (2) In the task of device status perception, the present invention extracts device status features layer by layer by adopting a deep forest algorithm with a multi-layer random forest structure, and optimizes the decision tree by combining a dynamic weight adjustment mechanism and a reinforcement learning strategy, thereby enhancing the model's ability to classify device status changes and abnormal states, and solving the problem of insufficient sensitivity to dynamic features in the device status classification process.

[0093] (3) In the task of equipment status perception, the present invention formulates a predictive maintenance plan based on the status classification results and risk level analysis, and optimizes the workload distribution by utilizing the equipment load timing characteristics, thereby solving the problem of overload and potential failures that cannot be effectively avoided during peak equipment operation periods.

[0094] (4) In the task of device status perception, the present invention obtains device operation logs and sensor data through real-time monitoring and historical record analysis, cleans and formats the data, and uses manual labeling to distinguish between normal, warning and fault states, thereby solving the problems of complex data sources and uneven data quality in device status perception. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 It is a loss function convergence curve diagram of the present invention. DETAILED DESCRIPTION

[0096] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0097] The equipment status perception system based on peak-shifting and dangerous situation awareness includes a data acquisition unit, a model training and optimization unit, a situation awareness unit, a peak-shifting control unit, a system interface and a user interaction unit.

[0098] The data collection unit is used to collect multi-source data and manually annotate the collected data. The annotated categories include: normal: indicating that the equipment is operating normally; warning: indicating that there are potential risks in the operation of the equipment; fault: indicating that a fault occurs in the operation of the equipment.

[0099] The training data of the equipment status perception model is collected from the operation logs and sensor records of various industrial equipment. The collection methods include real-time monitoring and historical record analysis to obtain comprehensive data on the equipment status. After data collection, it is converted into a format through an efficient data processing process and stored in a high-availability database in a structured JSON format.

[0100] The data attributes of the collected data include: equipment operating time Ra, equipment load percentage Da, ambient temperature Pa, ambient humidity Qa, equipment current Sa, equipment voltage Ta, equipment vibration frequency Ua, equipment noise level Va, preventive maintenance times Wa, and failure times Xa.

[0101] The model training and optimization unit processes the equipment status training data through the equipment status data feature dimensionality reduction model, adopts the autoencoder algorithm based on adaptive discarding as the equipment status data feature dimensionality reduction model, and dynamically calculates and updates the equipment status data feature similarity through Gaussian kernel function measurement and combined with gradient update terms, so that the main equipment status data features obtain weights in the dimensionality reduction process, effectively suppresses noise and improves the fidelity of equipment status data features. In the training stage of the autoencoder, the adaptive discarding probability and discarding feedback method are used to reduce redundant features, balance the simplification of equipment status data features and model performance, and improve the generalization ability and dimensionality reduction effect of the autoencoder model.

[0102] The model training and optimization unit relies on a variety of machine learning models to cooperate with each other to achieve equipment status prediction and anomaly detection, build an equipment status perception model based on training data, and achieve accurate characterization of equipment operating status and potential dangerous situations through multi-level feature processing.

[0103] The training process of the adaptive dropout-based autoencoder algorithm includes the following steps:

[0104] S101, in the initial stage of dimensionality reduction of device state data features, firstly, by calculating the similarity between the input device state data feature vectors, the relationship between each device state data feature is dynamically evaluated, and the correlation relationship between the device state data features is established by preliminarily calculating and updating the feature similarity of the input device state data, so as to focus on the main device state data features with precise weights in the subsequent steps, and suppress irrelevant device state data features, which can effectively reduce the interference of high-dimensional device state data noise on the dimensionality reduction process and improve the fidelity of the device state data feature representation. The device state data feature similarity is initially measured by a Gaussian kernel function, and a gradient adjustment term is used to realize a dynamic update mechanism. The calculation method is expressed as:

[0105]

[0106] In the formula, is the adjustment item in the similarity matrix, representing the The samples and The change of samples at the feature gradient level, To control the parameters of the gradient update weights, Set to 0.3, and Respectively and The characteristic gradient of each sample is used to characterize the changing trend of the local characteristics of the device status data. is the first sample index input into the autoencoder, is the sample index of the second input to the autoencoder.

[0107] Based on the above adjustment items, the similarity matrix is ​​updated and the calculation method is expressed as:

[0108]

[0109] In the formula, For the The samples and The similarity of the updated samples is is the similarity calculation function, For the The samples and The square of the Euclidean distance between samples is used to characterize the difference between the two in the feature space of the original device status data. is the L2 norm, is the first input to the autoencoder samples, is the first input to the autoencoder samples, The smoothing parameter is the Gaussian kernel control coefficient Set to 0.3.

[0110] S102, training the adaptive encoding and decoding network of the autoencoder, mapping the high-dimensional device status data features to a low-dimensional space and reconstructing them through the autoencoder structure, and using a weighted reconstruction error function based on similarity to enhance the network's attention to key device status data features, thereby retaining important information while reducing the dimensionality. During the training process of the autoencoder, the calculation method of the weighted reconstruction loss function is expressed as:

[0111]

[0112] In the formula, is the weighted reconstruction error of the autoencoder, is the number of samples input to the autoencoder for the current batch, For the The discard status of samples, For the The weighting coefficient of samples is used to strengthen the focus on important features. is the first input to the autoencoder samples, representing the The original device status data feature vector of samples, For the The device status data feature vector reconstructed by samples is For distance loss.

[0113] In order to balance the similarity and the distance between samples, the weighted coefficient is calculated based on the similarity matrix, and the calculation method is expressed as:

[0114]

[0115] In the formula, To determine the trade-off coefficient between similarity and distance, Set to 0.3.

[0116] S103, in order to enhance the modeling capability of potential nonlinear relationships in the device state data, the feature patterns of complex device state data are captured and amplified through nonlinear mapping and reweighting operations. The input device state data feature matrix is ​​subjected to nonlinear transformation, so as to better characterize the nonlinear structure of the original device state data in the new device state data feature space. The calculation method of the nonlinear mapping is expressed as:

[0117]

[0118] In the formula, For the The nonlinear characteristics after sample mapping, is the hyperbolic tangent function, which is used to simulate nonlinear mapping. is the weight matrix of the autoencoder, For the The noise intensity of each sample, is the bias term of the autoencoder.

[0119] In order to enhance the contribution of nonlinear features to reconstruction and dimensionality reduction, the distance loss is calculated as follows:

[0120]

[0121] In the formula, For the The feature weighting coefficient of samples in the nonlinear feature space, is the parameter that controls the sensitivity of class differentiation, Set to 0.2, For the The Euclidean distance between the nonlinear features of a sample and its true category, is the Euclidean distance calculation function.

[0122] The feature reweighting coefficient of equipment status data can amplify important features in nonlinear space, and the calculation method is expressed as:

[0123]

[0124] In the formula, is the similarity calculation function, For the The similarity measurement result between a sample and other samples is To control the parameters of the similarity influence range, Set to 0.1, For the The feature change gradient of a sample is used to reflect the importance of the feature in the update direction.

[0125] A non-stationary noise suppression method based on physical modeling is adopted to estimate the dynamic distribution of noise in the feature space of equipment status data through physical parameter modeling, and suppress the noise during dimensionality reduction, thereby improving the feature quality and model stability of the equipment status data after dimensionality reduction. Assuming that the noise intensity changes non-stationarily between samples, the calculation method is expressed as:

[0126]

[0127] In the formula, For the The noise intensity of each sample, is the noise baseline intensity, Set to 0.3, is the noise amplitude coefficient, which measures the range of noise fluctuations. Set to 0.1, The frequency coefficient of is the parameter of the noise change speed, Set to , For the The position index of a sample in the feature space.

[0128] S104, for redundant features in the training process, an adaptive discard probability is used to control the retention and discarding of device status data features. When some features have a low contribution or are highly similar to other features, the features are discarded with a higher probability, thereby simplifying the model and enhancing the effectiveness of the representation of device status data features. The discard operation is represented by a binary random variable, and the calculation method is expressed as:

[0129]

[0130] In the formula, For the The discard status of a sample, indicating whether the feature is retained.

[0131] In order to make the discard probability more accurately reflect the actual contribution of the device status data features in the current training round, the discard probability is calculated based on the original reconstruction error and the additional gain term. The calculation method is expressed as:

[0132]

[0133] In the formula, For the The probability of discarding a sample is is the adjustment parameter of the discard probability, For the The reconstruction error of samples is For the The gain term of samples;

[0134] The sample gain term is used to strengthen the focus on important features, and the calculation method is expressed as:

[0135]

[0136] In the formula, For the The gain term of samples, For the The weighting coefficient of samples, is the smoothing coefficient, is the L2 norm.

[0137] S105, adopt the method of discarding feedback to modify the discarding strategy according to the reconstruction situation after training. If discarding the device status data features leads to poor overall reconstruction effect, the discarding amount is reduced accordingly and the discarding probability is redistributed, so as to strike a balance between streamlining the device status data features and maintaining the model performance. This process is constrained by feedback loss, and the calculation method is expressed as:

[0138]

[0139] In the formula, is the feedback loss after discarding, For the The reconstruction error of samples is is the feedback weight, Set to 0.3, For the The discard status of a sample determines the impact of the discard strategy on model training.

[0140] S106, in order to avoid overfitting and accelerate convergence, the adaptive learning rate adjustment and optimal dimension reduction dimension selection method is adopted. The learning rate is adjusted in real time according to the changes in the reconstruction error and the discarding mechanism to avoid local convergence and accelerate model training. The optimal dimension reduction dimension is selected by using information gain and feature variance evaluation to ensure that the device status data after dimension reduction still retains the core information and has good discrimination. The learning rate update method is expressed as:

[0141]

[0142] In the formula, is the learning rate of the autoencoder at the current iteration, is the learning rate of the autoencoder in the previous iteration, Adjust the step size for the learning rate of the autoencoder, Set to 0.3, is the average reconstruction error of this round of training, is an additional dynamic coefficient used to fine-tune the learning rate in combination with the dropout probability. Set to 0.5, is the learning rate adjustment parameter, Set to 0.1.

[0143] In order to select the optimal dimension for dimensionality reduction, a comprehensive measurement method based on information gain and feature variance is used to select the optimal dimension for dimensionality reduction. The calculation method is expressed as:

[0144]

[0145] In the formula, is the optimal dimension reduction dimension, For the The information gain of the device status data feature is used to measure its identifiability. is the index of the device status data feature, is the characteristic dimension of the equipment status data before dimensionality reduction, For the The variance of the device status data feature is used to measure its discrimination in the sample distribution.

[0146] In order to measure the degree of complementarity between the characteristics of the equipment status data, the calculation method of feature information complementarity is used to measure the degree of complementarity between the characteristics of the equipment status data, so as to better decide which features should be retained. The calculation method is expressed as:

[0147]

[0148] In the formula, For the The first The information complementarity between the characteristics of device status data, For the Shannon entropy of the device status data features, For the Shannon entropy of the device status data features, For the The device status data characteristics and The joint entropy of the device status data features;

[0149] Use information complementarity to determine whether the The device status data characteristics and The device status data features are used to capture the core information of the device status data more comprehensively. For example, when the information complementarity between the features is greater than the preset threshold, the first The device status data characteristics and Device status data characteristics.

[0150] S107, repeat the above steps until a preset stop iteration condition is met. The preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times, which means that the model training is completed.

[0151] By calculating the convergence of the algorithm's loss function, the rationality of the number of iterations is verified. The experimental data are as follows: Figure 1 shown.

[0152] In order to verify the effectiveness of the proposed feature dimension reduction method, it is compared with the principal component analysis method. The comparative experimental data are shown in Table 1:

[0153] Table 1 Comparison data between the autoencoder algorithm based on adaptive discarding and the principal component analysis method

[0154]

[0155] The reduced-dimensional features are input into the device state perception model for device state recognition. The deep forest algorithm based on dynamic learning is used as the device state perception model. Through its classification performance, the collaboration between decision trees is optimized in combination with the dynamic learning process to improve the classification accuracy. The deep forest is a new algorithm that combines deep learning ideas and integrated learning methods. It aims to solve the learning problems of small samples and low-dimensional data. It extracts features layer by layer through the hierarchical structure of the cascaded random forest and gradually improves the expression ability of the model.

[0156] The training process of the deep forest algorithm based on dynamic learning includes the following steps:

[0157] S201, input the device status data after feature dimension reduction into the deep forest algorithm based on dynamic learning, establish a multi-layer deep forest structure, and assign initial weights to each random forest in the deep forest to solve the problem of how to effectively combine multi-layer information in the overall model and improve the model's deep processing ability to cope with nonlinear data. In the device status perception classification task, the device status data after feature dimension reduction is usually highly nonlinear. The weight allocation mechanism helps the model pay more attention to the key feature changes of different device states, thereby improving the classification accuracy of different device states. The weight calculation method of each layer of random forest is expressed as:

[0158]

[0159] In the formula, For the The weights of the random forest, is the index of the first random forest number, is the number of statistical items, including overall accuracy, class accuracy, kappa coefficient, etc. is the adjustment coefficient of error rate to weight, Set to 0.3, For the The error rate corresponding to each statistical item.

[0160] S202, dynamically adjust the partial derivatives of the weights of the random forest based on the loss function to solve the problem that each random forest is insensitive to error feedback during the training process, so that the model can adapt to changes in data distribution more quickly. The distribution of equipment status data often changes dynamically due to changes in operating conditions or environmental conditions. The deep forest model dynamically adjusts the partial derivatives of the random forest weights so that it can quickly respond to changes in the distribution of equipment status data. When the performance of certain equipment status features is affected by changes in operating conditions, the dynamic adjustment enables the model to strengthen the classification ability of key features in real time and reduce the occurrence of misclassification, which is expressed as:

[0161]

[0162] In the formula, For the The amount by which the random forest weights are adjusted, is the index of the second random forest number, is the learning rate of the random forest weights, Set to 0.3, is the number of random forests in the deep forest, For the The balance coefficient of the random forest loss is preset manually. Based on The loss function calculated by a random forest is is the symbol of partial derivative.

[0163] S203, the decision tree structure is self-optimized by using a reinforcement learning strategy to solve the problem of insufficient adaptation to novel and abnormal data under a static structure, so that the model can still maintain a high recognition ability when facing variable and complex data distribution. In the equipment status perception classification task, the operating status of the equipment may have unknown or abnormal categories. Deep Forest dynamically optimizes the decision tree structure through a reinforcement learning strategy, so that the model can still maintain high classification performance when facing novel or abnormal equipment status. When rare equipment failure modes appear in the equipment status data, the reinforcement strategy can help the decision tree quickly capture and adapt to the new mode, thereby improving the generalization ability of classification, which is expressed as:

[0164]

[0165] In the formula, is the performance function parameter before updating, is the updated performance function parameter, is the learning step size of the performance function, Set to 0.3, is the gradient of the performance function, is the performance function, It is the performance function parameter.

[0166] The performance function is based on an independent evaluation of each decision tree in the random forest and then a comprehensive score, and the calculation method is expressed as:

[0167]

[0168] In the formula, is the index of the number of decision trees, is the total number of decision trees in the random forest, For the The weight of a decision tree is preset manually. For the The classification accuracy of a decision tree under the current parameters.

[0169] S204, using a comprehensive method based on information gain and distribution characteristics to evaluate and screen features, solves the problem of how to focus on the most discriminative features in nonlinear situations, reduces noise interference and improves the discriminative power of the overall model. When predicting equipment failure status, high-variance vibration features may better reflect changes in equipment status, while low-variance features may be noise. By using a comprehensive method based on information gain and distribution characteristics to evaluate and screen features, it is ensured that the model focuses more accurately on core features, thereby improving the classification effect, which is expressed as:

[0170]

[0171] In the formula, For the Features, is the information gain function, For the dataset The entropy of is the dataset input to the deep forest, Features Value The probability of is the possible value of the feature, is the set of eigenvalues, is the information gain regularization coefficient, Set to 0.2, Features In the dataset The variance on .

[0172] According to the specific definition of variance, the information gain is corrected. The calculation method of variance is expressed as:

[0173]

[0174] In the formula, For the data points in the feature The value on is the index of the data point input to the deep forest, Features The mean of is the size of the dataset input to the deep forest.

[0175] S205, after completing the feature importance evaluation and selection, the adaptive feature transformation layer is further used to solve the problem of insufficient adaptation of feature representation when the data distribution changes. The input features are re-encoded more flexibly during the training process to enhance the generalization ability of the model in complex environments. When the equipment state switches to an abnormal working condition, the adaptive feature transformation layer reconstructs the key features through nonlinear activation, so that the model can better capture the significant features in the state transition, thereby improving the accuracy of abnormal state classification, which is expressed as:

[0176]

[0177] In the formula, is the adaptive feature transformation layer function, is the input feature vector of the deep forest, is the weight matrix of the adaptive feature transformation layer, is the bias vector of the adaptive feature transformation layer, is the nonlinear activation function of the adaptive feature transformation layer, and the activation function of the adaptive feature transformation layer adopts the ReLU activation function.

[0178] The feature transformation layer parameters are updated based on the loss function of feature importance and classification effect. The loss function calculation method of the feature transformation layer is expressed as:

[0179]

[0180] In the formula, is the loss function of the feature transformation layer, The number of samples input to the feature transformation layer for the current batch, The index of the sample input to the feature transformation layer for the current batch, is the cross entropy loss, For the The true labels of samples, For the model The predicted output of samples is is the regularization coefficient of the feature transformation layer, Set to 0.3, For The regularization function is L2 regularization.

[0181] Parameter update is performed according to the back propagation algorithm, and the calculation method is expressed as:

[0182]

[0183]

[0184] In the formula, is the weight matrix of the updated adaptive feature transformation layer, is the weight matrix of the adaptive feature transformation layer before updating, is the bias vector of the updated adaptive feature transformation layer, is the bias vector of the adaptive feature transformation layer before updating, is the learning rate of the feature transformation layer, is the gradient of the loss function of the feature transformation layer with respect to the weight matrix of the adaptive feature transformation layer, is the gradient of the loss function of the feature transformation layer with respect to the bias vector of the adaptive feature transformation layer.

[0185] S206, in the inter-layer feedback mechanism stage, by allowing the high-level output and feedback signals of the deep forest to act on the input of the next layer, the problem of information fragmentation between different layers is solved, and the coordination between the layers and the ability to capture complex patterns are strengthened. In the classification of device status perception, different layers may capture different characteristics of the device status. In order to solve the problem of information fragmentation between layers, the deep forest uses high-level output and feedback signals to act on the input of the next layer to strengthen the collaboration between layers. In the process of the device status gradually deteriorating from normal operation to failure, the inter-layer feedback transfers the gradual features extracted by the upper layer to the lower layer, ensuring that the model has a more accurate perception of the step-by-step changes in the device status, which is expressed as:

[0186]

[0187] In the formula, For the The input of a random forest is For the The output of a random forest, For the The regulation signal fed back by the random forest is The nonlinear mapping result of a random forest is is the activation function of random forest.

[0188] The specific inter-layer interaction is realized by the network layer that fuses the output and feedback signals. In the task of classifying the operating status of industrial equipment, the classification accuracy of normal operation, fault and abnormal status of equipment can be improved, and reliable technical support can be provided for equipment maintenance and monitoring. The activation function calculation method of random forest is expressed as:

[0189]

[0190] In the formula, is the Sigmoid activation function, For the The bias of a random forest.

[0191] The situation awareness unit applies the trained model to actual data input to perform classification, anomaly detection, or fault prediction for equipment status awareness.

[0192] Use the trained feature dimensionality reduction model to process the input device status data, pass the reduced features to the trained deep forest model, and rely on its multi-level random forest structure to achieve classification. Output the classification label or anomaly detection result of the device status prediction, and judge the risk level based on the classification probability. For example, the categories include: normal: indicates that the device is operating normally; warning: indicates that there is a potential risk in the operation of the device; fault: indicates that the device has a fault.

[0193] The peak-shifting control unit optimizes the equipment's operating time and workload through a peak-shifting strategy based on the timing characteristics of the equipment load and the prediction results of the operating status, thereby reducing the risk of equipment overload and failure during peak hours.

[0194] Based on the equipment status perception results, further predictive maintenance of the equipment is carried out, combining the real-time operation data of the equipment, historical fault records and the prediction results of the current model to plan and execute maintenance strategies.

[0195] First, based on the model's prediction output, i.e., equipment status classification (normal, warning, fault) and the corresponding risk assessment level, the predictive maintenance strategy should take different response measures for different risk levels.

[0196] For equipment whose status is classified as "normal", the goal of predictive maintenance is to ensure that the equipment continues to operate stably, while detecting whether there are signs of potential problems through data monitoring. At this stage, maintenance activities mainly focus on routine inspections and preventive maintenance, such as regular cleaning, lubrication, and replacement of consumables, to avoid unforeseen failures after long-term operation of the equipment. Based on the historical status data of the equipment, a maintenance cycle model is established, and the maintenance plan is adjusted according to the actual operation of the equipment.

[0197] For equipment in the "warning" category, predictive maintenance is more proactive. In this case, although the equipment has not completely failed, it has already shown potential risks and may fail in the near future. For this type of equipment, maintenance measures need to be more timely and accurate. First, through in-depth analysis of the equipment's status characteristics (such as temperature, vibration, current, etc.), combined with the equipment's operating environment and load conditions, predict factors that may cause failures, and take targeted measures to intervene.

[0198] For equipment predicted to be "faulty", immediate maintenance measures are taken. The equipment is already in a high-risk state and needs to be shut down as soon as possible for troubleshooting and repair. Through an intelligent predictive maintenance system, the equipment status can be continuously monitored and analyzed before a fault occurs, and the time, type and specific location of the fault can be accurately predicted.

[0199] The system interface and user interaction unit provide a user-friendly interactive interface, support status visualization, alarm notification and strategy adjustment. Through the graphical interface, users can intuitively view the equipment operation status, dangerous situation warning and peak control suggestions. The system interface and user interaction unit supports seamless integration with other systems, provides data export, API access and other functions, and facilitates users to apply the system in multiple scenarios. The interactive module also supports a feedback mechanism, allowing users to evaluate and modify the control strategy provided by the system, thereby further improving the intelligence level of the system.

[0200] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention.

Claims

1. The equipment status perception system based on peak shifting and dangerous situation perception is characterized by: include: The data collection unit is used to collect multi-source data and manually annotate the collected data. The annotated categories include: normal: indicating that the equipment is operating normally; warning: indicating that there are potential risks in the operation of the equipment; fault: indicating that the equipment has a fault; The model training and optimization unit processes the equipment state training data through the equipment state data feature dimensionality reduction model, and adopts the autoencoder algorithm based on adaptive discarding as the equipment state data feature dimensionality reduction model; The reduced-dimensional features are input into the device state perception model for device state recognition. The deep forest algorithm based on dynamic learning is used as the device state perception model to improve the classification accuracy. The hierarchical structure of the cascaded random forest is used to extract features layer by layer to gradually improve the model's expression ability. Situation awareness unit, which applies the trained model to actual data input to perform classification, anomaly detection or fault prediction for equipment status awareness; The peak-shifting control unit optimizes the equipment's operating time and workload through peak-shifting strategies based on the equipment load's timing characteristics and operating status prediction results, thus reducing the risk of equipment overload and failure during peak hours. The system interface and user interaction unit provide a user-friendly interactive interface, support status visualization, alarm notification and strategy adjustment. Through the graphical interface, users can intuitively view the equipment operation status, dangerous situation warnings and peak-shifting control suggestions.

2. The equipment status perception system based on peak shifting and dangerous situation perception according to claim 1 is characterized in that: The data attributes of the collected data include: equipment operating time Ra, equipment load percentage Da, ambient temperature Pa, ambient humidity Qa, equipment current Sa, equipment voltage Ta, equipment vibration frequency Ua, equipment noise level Va, preventive maintenance times Wa, and failure times Xa.

3. The equipment status perception system based on peak shifting and dangerous situation perception according to claim 1 is characterized in that: The training process of the self-encoder algorithm based on adaptive discarding includes the following steps: S101, in the initial stage of dimensionality reduction of device state data features, firstly, by calculating the similarity between the input device state data feature vectors, the relationship between each device state data feature is dynamically evaluated, and the correlation relationship between the device state data features is established by preliminarily calculating and updating the feature similarity of the input device state data. In the subsequent steps, the main device state data features are focused with precise weights, and irrelevant device state data features are suppressed. The device state data feature similarity is initially measured by a Gaussian kernel function, and a gradient adjustment term is used to implement a dynamic update mechanism. The calculation method is expressed as: In the formula, is the adjustment item in the similarity matrix, representing the The samples and The change of samples at the feature gradient level, To control the parameters of the gradient update weights, and Respectively and The characteristic gradient of each sample is used to characterize the changing trend of the local characteristics of the device status data. is the first sample index input into the autoencoder, is the sample index of the second input into the autoencoder; Based on the above adjustment items, the similarity matrix is ​​updated and the calculation method is expressed as: In the formula, For the The samples and The similarity of the updated samples is is the similarity calculation function, For the The samples and The square of the Euclidean distance between samples is used to characterize the difference between the two in the feature space of the original device status data. is the L2 norm, is the first input to the autoencoder samples, is the first input to the autoencoder samples, The smoothing parameter is the Gaussian kernel control coefficient; S102, training the adaptive encoding and decoding network of the autoencoder, mapping the high-dimensional device status data features to a low-dimensional space and reconstructing them through the autoencoder structure, using a weighted reconstruction error function based on similarity to enhance the network's attention to key device status data features, retaining important information while reducing the dimensionality. During the training process of the autoencoder, the calculation method of the weighted reconstruction loss function is expressed as: In the formula, is the weighted reconstruction error of the autoencoder, is the number of samples input to the autoencoder for the current batch, For the The discard status of samples, For the The weighting coefficient of samples is used to strengthen the focus on important features. is the first input to the autoencoder samples, representing the The original device status data feature vector of samples, For the The device status data feature vector reconstructed by samples is For distance loss; In order to balance the similarity and the distance between samples, the weighted coefficient is calculated based on the similarity matrix, and the calculation method is expressed as: In the formula, To determine the trade-off coefficient between similarity and distance; S103, through nonlinear mapping and re-weighting operations, the characteristic pattern of complex device status data is captured and amplified, and a nonlinear transformation is performed on the input device status data characteristic matrix. The calculation method of the nonlinear mapping is expressed as: In the formula, For the The nonlinear characteristics after sample mapping, is the hyperbolic tangent function, which is used to simulate nonlinear mapping. is the weight matrix of the autoencoder, For the The noise intensity of each sample, is the bias term of the autoencoder; In order to enhance the contribution of nonlinear features to reconstruction and dimensionality reduction, the distance loss is calculated as follows: In the formula, For the The feature weighting coefficient of samples in the nonlinear feature space, is a parameter that controls the sensitivity of class distinction. For the The Euclidean distance between the nonlinear features of a sample and its true category, is the Euclidean distance calculation function; The feature reweighting coefficient of equipment status data can amplify important features in nonlinear space, and the calculation method is expressed as: In the formula, is the similarity calculation function, For the The similarity measurement result between a sample and other samples is To control the parameters of the similarity influence range, For the The feature change gradient of a sample is used to reflect the importance of the feature in the update direction; The non-stationary noise suppression method based on physical modeling is adopted to estimate the dynamic distribution of noise in the feature space of equipment status data through physical parameter modeling, and suppress the noise during dimensionality reduction. The calculation method is expressed as: In the formula, For the The noise intensity of each sample, is the noise baseline intensity, is the noise amplitude coefficient, which measures the range of noise fluctuations. The frequency coefficient of is the parameter of the noise change speed, For the The position index of each sample in the feature space; S104, for redundant features in the training process, an adaptive discard probability is used to control the retention and discarding of device status data features. The discard operation is represented by a binary random variable, and the calculation method is expressed as: In the formula, For the The discard status of a sample, which indicates whether the feature is retained; The discard probability is calculated based on the original reconstruction error and the additional gain term, and the calculation method is expressed as: In the formula, For the The probability of discarding a sample is is the adjustment parameter of the discard probability, For the The reconstruction error of samples is For the The gain term of samples; The sample gain term is used to strengthen the focus on important features, and the calculation method is expressed as: In the formula, For the The gain term of samples, For the The weighting coefficient of samples, is the smoothing coefficient, is the L2 norm; S105, adopting the discard feedback method to modify the discard strategy according to the reconstruction situation after the training is completed, and the calculation method is expressed as: In the formula, is the feedback loss after discarding, For the The reconstruction error of samples is the feedback weight, For the The discard status of each sample determines the impact of the discard strategy on model training; S106 adopts the method of adaptive learning rate adjustment and optimal dimension reduction dimension selection. By adjusting the learning rate in real time according to the changes in reconstruction error and discarding mechanism, local convergence is avoided and model training is accelerated. The optimal dimension reduction dimension is selected by using information gain and feature variance evaluation. The learning rate update method is expressed as: In the formula, is the learning rate of the autoencoder at the current iteration, is the learning rate of the autoencoder in the previous iteration, Adjust the step size for the learning rate of the autoencoder, is the average reconstruction error of this round of training, is an additional dynamic coefficient used to fine-tune the learning rate in combination with the drop probability. Set to 0.5, Tuning parameters for the learning rate A comprehensive measurement method based on information gain and feature variance is used to select the optimal dimension reduction dimension. The calculation method is expressed as: In the formula, is the optimal dimension reduction dimension, For the The information gain of the device status data feature is used to measure its identifiability. is the index of the device status data feature, is the characteristic dimension of the equipment status data before dimensionality reduction, For the The variance of the device status data feature is used to measure its discrimination in the sample distribution; The calculation method of feature information complementarity is used to measure the degree of complementarity between the features of equipment status data. The calculation method is expressed as: In the formula, For the The first The information complementarity between the characteristics of device status data, For the Shannon entropy of the device status data features, For the Shannon entropy of the device status data features, For the The device status data characteristics and The joint entropy of the device status data features; S107, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

4. The equipment status perception system based on peak shifting and dangerous situation perception according to claim 1 is characterized in that: The training process of the deep forest algorithm based on dynamic learning includes the following steps: S201, input the device status data after feature dimension reduction into the deep forest algorithm based on dynamic learning, establish a multi-layer deep forest structure, and assign initial weights to each random forest in the deep forest to improve the model's deep processing capability for nonlinear data. The weight calculation method for each layer of random forest is expressed as: In the formula, For the The weights of the random forest, is the index of the first random forest number, is the number of statistical items, is the adjustment coefficient of error rate to weight, For the The error rate corresponding to each statistical item; S202, dynamically adjust the partial derivative of the random forest weight based on the loss function. The deep forest model dynamically adjusts the partial derivative of the random forest weight, which is expressed as: In the formula, For the The amount by which the random forest weights are adjusted, is the index of the second random forest number, is the learning rate of the random forest weights, is the number of random forests in the deep forest, For the The balance coefficient of the random forest loss is preset manually. Based on The loss function calculated by a random forest is is the symbol of partial derivative; S203, the decision tree structure is self-optimized using a reinforcement learning strategy, so that the model can still maintain a high recognition ability when facing variable and complex data distribution. Deep Forest dynamically optimizes the decision tree structure through a reinforcement learning strategy, which is expressed as: In the formula, is the performance function parameter before updating, is the updated performance function parameter, is the learning step size of the performance function, is the gradient of the performance function, is the performance function, is the performance function parameter; The performance function is based on an independent evaluation of each decision tree in the random forest and then a comprehensive score, and the calculation method is expressed as: In the formula, is the index of the number of decision trees, is the total number of decision trees in the random forest, For the The weight of a decision tree is preset by human. For the The classification accuracy of a decision tree under the current parameters; S204, a comprehensive method based on information gain and distribution characteristics is used to evaluate and screen features, reduce noise interference and improve the discriminative power of the overall model, which is expressed as: In the formula, For the Features, is the information gain function, For the dataset The entropy of is the dataset input to the deep forest, Features Value The probability of is the possible value of the feature, Values ​​is the feature value set, is the information gain regularization coefficient, Features In the dataset The variance on ; According to the specific definition of variance, the information gain is corrected. The calculation method of variance is expressed as: In the formula, For the data points in the feature The value of is the index of the data point input to the deep forest, Features The mean of is the size of the dataset input to the deep forest; S205, after completing the feature importance evaluation and selection, an adaptive feature transformation layer is used to re-encode the input features more flexibly during the training process to enhance the generalization ability of the model in complex environments, which is expressed as: In the formula, is the adaptive feature transformation layer function, is the input feature vector of the deep forest, is the weight matrix of the adaptive feature transformation layer, is the bias vector of the adaptive feature transformation layer, is the nonlinear activation function of the adaptive feature transformation layer, and the activation function of the adaptive feature transformation layer adopts the ReLU activation function; The feature transformation layer parameters are updated based on the loss function of feature importance and classification effect. The loss function calculation method of the feature transformation layer is expressed as: In the formula, is the loss function of the feature transformation layer, The number of samples input to the feature transformation layer for the current batch, The index of the sample input to the feature transformation layer for the current batch, is the cross entropy loss, For the The true labels of samples, For the model The predicted output of samples is is the regularization coefficient of the feature transformation layer, For Regularization function, specifically L2 regularization; Parameter update is performed according to the back propagation algorithm, and the calculation method is expressed as: , In the formula, is the weight matrix of the updated adaptive feature transformation layer, is the weight matrix of the adaptive feature transformation layer before updating, is the bias vector of the updated adaptive feature transformation layer, is the bias vector of the adaptive feature transformation layer before updating, is the learning rate of the feature transformation layer, is the gradient of the loss function of the feature transformation layer with respect to the weight matrix of the adaptive feature transformation layer, is the gradient of the loss function of the feature transformation layer with respect to the bias vector of the adaptive feature transformation layer; S206, in the inter-layer feedback mechanism stage, strengthen the coordination between layers and the ability to capture complex patterns, expressed as: In the formula, For the The input of a random forest is For the The output of a random forest, For the The regulation signal fed back by the random forest is The nonlinear mapping result of a random forest is is the activation function of random forest; The specific inter-layer interaction is realized based on the network layer that fuses the output and feedback signals, providing reliable technical support for equipment maintenance and monitoring. The activation function calculation method of random forest is expressed as: In the formula, is the Sigmoid activation function, For the The bias of a random forest.

Images