A method for predicting the service life of electromechanical equipment

Abstract

The present invention belongs to the technical field of electromechanical equipment management and maintenance, and particularly relates to a method for predicting the life of electromechanical equipment. It mainly includes the following steps: determining the expected life correlation attribute sequence according to the type of electromechanical equipment; determining the data to be analyzed and the initial reference data; screening samples from the samples to establish a sample database, selecting a suitable data period, using a sliding time window to intercept the corresponding period monitoring data for similarity analysis to determine the similarity and extract effective data, calculating the period change rate deviation to optimize the samples and establish an effective reference database; establishing a prediction weight and prediction based on the time series similarity of the sample data. The electromechanical equipment life prediction method of the present application is mainly used to provide a method applicable to new equipment, which can better consider the correlation between historical measurement data and the life cycle, manifest the relationship between the operating states of different life stages of electromechanical equipment and the final expected life, and better analyze and predict the expected remaining life of electromechanical equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of electromechanical equipment management and maintenance, and particularly relates to a method for predicting the life of electromechanical equipment. Background Art

[0002] The prediction of the life of electromechanical equipment refers to predicting the remaining life period of the equipment based on the current operating parameters or attributes of the equipment and combining the operating data of similar equipment, so as to facilitate the advance planning and design of the replacement and renewal process of electromechanical equipment, ensure the continuity of equipment functions, and avoid unnecessary downtime and other problems. Currently, the prediction of the life of electromechanical equipment mainly analyzes and predicts based on the current attributes or parameters of the equipment, and rarely considers the associated impact of historical cycle data on the equipment life. This is because the expected life of most current equipment is often strongly associated with the available state of a small number of core structures. These core structures often have a relatively fixed expected life and have a very clear predictable state change process when damaged. Therefore, by paying attention to the current attributes or parameters and then determining the state of the core structure, the expected remaining life of the equipment can be estimated. However, with the application of flexible equipment and the increasingly mature equipment maintenance technology and capabilities, while the life cycle of future electromechanical equipment is further extended, its expected life will also be further associated with the specific operating processes at different stages within its entire life cycle. In this case, the traditional life prediction method cannot well reflect the relevance of the corresponding life cycle and cannot further improve the effectiveness of the equipment life prediction result. Summary of the Invention

[0003] The purpose of the present invention is to propose a method that takes into account the historical measurement data within the equipment life cycle and can better predict the life of new electromechanical equipment in combination with actual needs.

[0004] To achieve the above purpose, the present invention adopts the following technical solutions.

[0005] A method for predicting the life of electromechanical equipment in the present application mainly includes the following steps:

[0006] Stp1. Determine the expected life correlation attribute sequence according to the type of electromechanical equipment ; Based on the foregoing correlation attribute sequence, determine the existing monitoring data containing each attribute in the correlation attribute sequence from the monitoring data of the electromechanical equipment to be analyzed as the data to be analyzed; Select the historical monitoring data covering the corresponding cycle or stage from the historical monitoring data of existing samples of the same type of electromechanical equipment as the initial reference data;

[0007] Stp2. Preferably, establish a database with the selected data from the existing samples to form a sample database; select an appropriate data period based on the data collection period characteristics of the sample to be analyzed and the same type of equipment , ; where refers to the sampling period of any attribute element . To ensure the validity of the sample data of the attribute element is the minimum period

[0008] For each attribute element in the attribute sequence of the sample , use a sliding time window to intercept the corresponding periodic monitoring data. Based on the periodic monitoring data intercepted by the sliding time window , generate a periodic monitoring data curve for the intercepted attribute element . Conduct a similarity analysis with the periodic monitoring data curve of the corresponding attribute element of the sample to be analyzed, determine the time window with the maximum similarity of the periodic monitoring data for each attribute element , and extract the periodic monitoring data within this time window as valid data to retain, obtaining the sequence of attribute values of the attribute element of the sample that is closest within the corresponding period . ; Plot the parameter change curve within the time window to determine the periodic change rate of the attribute element of the sample ;

[0009] where ;

[0010] Similarly, obtain the periodic change rate of the parameters of the sample to be analyzed ;

[0011] Calculate the historical data to calculate the periodic change rate deviation between the sample to be analyzed and each existing sample ; where represents the sampling data of the th type of parameter of the sample in the th sampling period; is the sampling period of the attribute element

[0012] Arrange the existing samples Sort, preferably select several existing samples, form the effective data sets of each sample from the corresponding effective data, and organize each sample to establish an effective reference database containing all samples;

[0013] Stp3. Establish expressions for prediction weights and prediction lifetimes based on the temporal similarity of sample data, including obtaining a matrix of samples to be analyzed expressed as a multi-dimensional vector from the data to be analyzed and the matrix of samples to be analyzed ;

[0014] ;

[0015] ;

[0016] Based on the sample in the previous step of the attribute elements of the periodic change rate to determine the periodic change matrix of the sample to be analyzed and the matrix of samples to be analyzed ; where:

[0017]

[0018] ;

[0019] Since the change of the equipment life cycle is often accompanied by the continuous change of the equipment performance index, the attribute data at each time node cannot well reflect the temporal change trend of the corresponding attribute characteristics. According to experience, when the life stage of a certain attribute data is closer to the current life stage of the product to be analyzed, the life expectancy of the equipment under the current data is also closer to the life expectancy of the product to be analyzed; to reflect this kind of influence, a stage matching index is established to characterize the matching degree of the life stage between the sample and the product to be analyzed , , the stage matching index characterizes the proximity of the life cycle node of this data to the current life cycle node of the equipment to be analyzed. The closer it is, the larger the stage matching index is, and vice versa. The faster the equipment life cycle changes, the larger the magnitude of the stage matching index is, and vice versa;

[0020] Based on the principle of similarity measurement, the similarity between the periodic change matrix of the sample to be analyzed and the reference sample matrix can be expressed as ; dimensionless processing is performed on the similarity to obtain a reference sample relative to the sample to be analyzed Degree of influence ;

[0021] To determine the influence weights of multiple reference samples on the sample to be analyzed in life prediction, while ensuring that the weight value range is , given a positive real number , obtain the reference The weight expression form of the sample relative to the sample to be analyzed ; After normalization, its weight expression is ; Wherein ; is the total number of reference samples;

[0022] Then the predicted life of the sample to be analyzed based on the reference sample can be expressed as:

[0023]

[0024] Wherein, represents the time node The expected remaining life of the sample to be analyzed at the moment , represents the time node The remaining life of the reference sample at the moment .

[0025] For further improvement or specific implementation of the aforementioned electromechanical equipment life prediction method, the data to be analyzed and the initial reference data include the historical monitoring data of each associated attribute in the associated attribute sequence The original monitoring data includes multiple groups of original monitoring data under different working conditions of the normal operation of the electromechanical equipment.

[0026] For further improvement or specific implementation of the aforementioned electromechanical equipment life prediction method, in the step Stp3, since the acquisition period of each attribute element is different, and the total amount of attribute data is also different, in the actual implementation process, the matrix parameters are optimized through methods such as zero-value filling and normalization processing to facilitate the operation and processing.

[0027] For further improvement or specific implementation of the aforementioned electromechanical equipment life prediction method, in the step Stp1, to ensure the effectiveness of each attribute element in the associated attribute sequence , and to ensure that the selected attribute elements can characterize the expected life of the electromechanical equipment, each element in the associated attribute is determined based on the following method:

[0028] A1. First, obtain the original monitoring data and perform necessary preprocessing, including deletion and revision of error data, normalization processing of the data. The original monitoring data includes multiple groups of original monitoring data under different working conditions of the normal operation of the electromechanical equipment;

[0029] After performing necessary preprocessing on the original data, it is segmented to obtain a training data set and a validation data set;

[0030] A2. Establish a multi-layer directed feature extraction model composed of multiple restricted Boltzmann machine network models (RBMs), where the restricted Boltzmann machine network model includes a visible layer and a hidden layer . The visible layer serves as the input side to obtain sample data, and the hidden layer serves as the output side to extract the main features of the sample and output them in the form of a probability distribution;

[0031] To optimize the data feature extraction effect, enhance data integrity, and avoid falling into local optima, the global weight change and local weight changes are used to control the weight changes of the hidden layer neurons, and the algorithm is optimized by increasing the number of hidden layers and the number of hidden layer neurons. Specifically:

[0032] The condition for increasing the number of hidden layer neurons can be expressed as: ;

[0033] The condition for increasing the number of hidden layers can be expressed as:

[0034] where ; ;

[0035] is the weight vector of the neuron of the hidden layer after iterations; refers to the global weight change of the neuron for the th training sample; is the total number of samples, is the number of samples that increased the global weight change in the previous iteration when ; where is the threshold function for the maximum global weight change to ensure that each hidden layer neuron can correspond to a training sample after increasing the number of hidden layer neurons,

[0036]

[0037] is the maximum number of iterations, represents the slope of the variable threshold function curve at the corresponding position, is the maximum value of the threshold function, is the minimum value of the threshold function, represents an energy function, is a preset control threshold; represents the number of hidden layers, represents the maximum number of current hidden layers;

[0038] During the training process, analyze the weight changes of the neurons in the hidden layer If the condition for increasing the number of neurons in the hidden layer is met, add a new neuron based on this neuron and reset the parameters of the two neurons to zero; at the same time, analyze whether the hidden layer meets the condition for increasing the number of hidden layers. If it meets the condition, copy this hidden layer and keep the original parameters unchanged;

[0039] A3. Based on the foregoing steps, establish and optimize a multi-layer directed feature extraction model, input the training data into the model for training, optimize and train the model using the validation data set to obtain the final model, extract the data change characteristics of the sample health state change process through the model, and screen through the correlation between these feature attributes and the correlation features of the electromechanical equipment health state to determine the preferred correlation attribute sequence. Description of the Drawings

[0040] Figure 1 is a schematic flowchart of the electromechanical equipment life prediction method in the embodiment;

[0041] Figure 2 is a schematic diagram of a multi-layer directed feature extraction model based on a multi-RBN network. Detailed Embodiments

[0042] The electromechanical equipment life prediction method of the present application is mainly used to provide a method applicable to new equipment, which can better consider the correlation between historical measurement data and the life cycle, manifest the relationship between the operating states of electromechanical equipment in different life stages and the final expected life, and better analyze and predict the expected remaining life of electromechanical equipment.

[0043] The following describes the present invention in detail with reference to specific embodiments.

[0044] As Figure 1 shown, the electromechanical equipment life prediction method of the present application mainly includes the following steps:

[0045] Step Stp1. Necessary data collection work;

[0046] Determine the expected life correlation attribute sequence according to the type of electromechanical equipment ; Based on the foregoing correlation attribute sequence, determine the existing monitoring data containing each attribute in the correlation attribute sequence from the monitoring data of the electromechanical equipment to be analyzed as the data to be analyzed;

[0047] Select historical monitoring data covering the corresponding period or stage from the historical monitoring data of existing electromechanical equipment samples of the same type as the initial reference data; The historical monitoring data of each associated attribute in the associated attribute sequence is included in the data to be analyzed and the initial reference data;

[0048] In the historical monitoring data of each associated attribute in the associated attribute sequence; The historical monitoring data;

[0049] Specifically, in this embodiment, to ensure the effectiveness of each attribute element in the associated attribute sequence and ensure that the selected attribute elements can characterize the expected life of the electromechanical equipment, the present application determines each element in the associated attribute based on the following method: To ensure the effectiveness of each attribute element in the associated attribute sequence and ensure that the selected attribute elements can characterize the expected life of the electromechanical equipment, the present application determines each element in the associated attribute based on the following method:

[0050] A1. First, obtain the original monitoring data and perform necessary preprocessing, including deleting and revising error data and normalizing the data. The original monitoring data includes multiple groups of original monitoring data under different working conditions of the normal operation of the electromechanical equipment;

[0051] After performing necessary preprocessing on the original data, it is segmented to obtain a training data set and a validation data set;

[0052] A2. As shown in Figure 2 , a multi-layer directed feature extraction model composed of multiple restricted Boltzmann machine network models (RBM) is established. The restricted Boltzmann machine network model includes a visible layer and a hidden layer . The visible layer obtains sample data as the input side, and the hidden layer extracts the main features of the sample and outputs them in the form of a probability distribution;

[0053] To optimize the data feature extraction effect, enhance the integrity of the data and avoid falling into local optima, and retain the relevance between the attribute parameters and the equipment life within the entire life cycle of the equipment, the global weight change amount and the local weight change are used to control the weight change of the hidden layer neurons, and the algorithm model is optimized by increasing the number of hidden layers and the number of hidden layer neurons. Specifically:

[0054] The condition for increasing the number of hidden layer neurons can be expressed as: ;

[0055] The condition for increasing the number of hidden layers can be expressed as:

[0056] where ; ;

[0057] is the neuron of the hidden layer after iterations The weight vector; Refers to the neuron For the Global weight change of the th training sample ; Is the total number of samples, During the previous iteration, the global weight change The number of samples that increase; where To ensure that after the number of hidden layer neurons increases, each hidden layer neuron can correspond to a training sample with the maximum global weight change The threshold function of Can be expressed as:

[0058]

[0059] Is the maximum number of iterations, Represents the slope at the corresponding position of the variable threshold function curve, Is the maximum value of the threshold function, Is the minimum value of the threshold function, Represents the energy function, Is the preset control threshold; Represents the number of hidden layers, Represents the maximum number of layers of the current hidden layer;

[0060] During the training process, analyze the weight changes of the hidden layer neurons If the condition for increasing the number of hidden layer neurons is met, a new neuron is added on the basis of this neuron, and the parameters of the two neurons are reset to zero; at the same time, analyze whether the hidden layer Meets the condition for increasing the number of hidden layer levels. If it meets, copy the hidden layer and keep the original parameters unchanged;

[0061] A3. Based on the above steps, establish and optimize a multi-layer directed feature extraction model, input the training data into the model for training, optimize and train the model using the validation data set to obtain the final model, and extract the data change characteristics of the sample health state change process through the model. Screen through the correlation between these feature attributes and the correlation features of the electromechanical equipment health state to determine the preferred correlation attribute sequence;

[0062] Based on the above correlation attribute sequence, determine the existing monitoring data containing each attribute in the correlation attribute sequence from the monitoring data of the electromechanical equipment to be analyzed As the data to be analyzed;

[0063] Select the historical monitoring data covering the corresponding period or stage from the historical monitoring data of the existing same-type electromechanical equipment samples As the initial reference data;

[0064] The data to be analyzed and the initial reference data include the historical monitoring data of each associated attribute in the associated attribute sequence. ;

[0065] Select a suitable data period based on the data acquisition cycle characteristics of the sample to be analyzed and the same type of equipment , ; where refers to the sampling period of any attribute element , To ensure the validity of the sample data of the attribute element The minimum period;

[0066] For the sample Attribute sequence Each attribute element in , use a sliding time window Intercept the corresponding periodic monitoring data, and based on the sliding time window The intercepted attribute element Generate a periodic monitoring data curve from the monitoring data, and perform similarity analysis with the periodic monitoring data curve of the corresponding attribute element of the sample to be analyzed to determine the time window with the largest similarity of the periodic monitoring data of each attribute element ;

[0067] Considering that the attribute element data has time scalability, and the attribute data may have different continuous periods at a certain stage in different devices. Therefore, based on the dynamic time warping (DTW) method, the data intercepted by the time window Can be used to form a data curve, and whether the corresponding data has a close life cycle can be determined by whether there is a high temporal correspondence between the data curves, and then the most matching time window can be determined.

[0068] Extract the periodic monitoring data within the time window as valid data and retain it to obtain the sample The attribute element of The sequence of attribute values closest within the corresponding period , ; Plot the parameter change curve within the time window To determine the periodic change rate Of the attribute element of the sample ; ;

[0069] where ;

[0070] Similarly, obtain the parameter periodic change rate Of the sample to be analyzed ;

[0071] Calculate historical data and the samples to be analyzed With each existing sample The deviation of the periodic change rate ; Wherein Represents the sample Of the Class parameters at the Sampling data of the nth acquisition cycle; Is the attribute element Sampling period of;

[0072] Sort the existing samples according to the principle of the deviation of the periodic change rate from small to large, select several existing samples, form the effective data sets of each sample from the corresponding effective data, and organize each sample to establish an effective reference database containing all samples; Sort the existing samples according to the principle of the deviation of the periodic change rate from small to large, select several existing samples, form the effective data sets of each sample from the corresponding effective data, and organize each sample to establish an effective reference database containing all samples;

[0073] Combined with the previous step, using the slope of the line connecting two consecutive data at two adjacent time nodes, the data change trend within the detection period can be determined. Using the slope matching degree between the product to be analyzed and the sample data, the matching degree of the life characteristics of the two can be determined, and then the consistency between the life changes represented by the data of different samples and the product to be analyzed can be analyzed, and using this as the weight, the expected life of the product to be analyzed can be predicted through the actual life of the sample;

[0074] Specifically, based on the data to be analyzed, obtain the sample matrix to be analyzed expressed by a multi-dimensional vector And the sample matrix to be analyzed ;

[0075] ;

[0076] ;

[0077] Among them, since the acquisition periods of each attribute element Are different, and the total amount of attribute data is also different. In the actual implementation process, the matrix parameters can be optimized through methods such as zero-value filling and normalization processing to facilitate operation and processing;

[0078] Based on the previous step of the sample Of the attribute element The periodic change rate of Determine the periodic change matrix of the sample to be analyzed And the sample matrix to be analyzed ; Wherein:

[0079]

[0080] ;

[0081] Since the changes in the equipment life cycle are often accompanied by continuous changes in the equipment performance indicators, the attribute data at each time node cannot well reflect the temporal change trend of the corresponding attribute characteristics. According to experience, when the life stage of a certain attribute data is closer to the current life stage of the product to be analyzed, the life expectancy of the equipment under the current data is also closer to the life expectancy of the product to be analyzed. To reflect this influence, a stage matching index is established to characterize the matching degree of the life stages of the sample and the product to be analyzed. , , the stage matching index characterizes the closeness of the life cycle node of this data to the current life cycle node of the equipment to be analyzed. The closer they are, the larger the stage matching index ; conversely, the smaller it is. The faster the equipment life cycle changes, the larger the magnitude of the stage matching index ; conversely, the smaller it is.

[0082] Based on the similarity measurement principle, the period change matrix of the sample to be analyzed and the reference sample matrix The similarity can be expressed as ; for the similarity Perform dimensionless processing to obtain the influence degree of the reference sample relative to the sample to be analyzed . ;

[0083] To determine the influence weights of multiple reference samples on the sample to be analyzed in life prediction and ensure that the weight value range is , given a positive real number , obtain the weight expression form of the reference sample relative to the sample to be analyzed ; after normalization, its weight expression is ; where is the total number of reference samples;

[0084] Then the predicted life of the sample to be analyzed based on the reference sample can be expressed as:

[0085]

[0086] Among them, represents the predicted remaining life of the sample to be analyzed at the time node , represents the remaining life of the reference sample at the time node .

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting the protection scope of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for predicting the service life of an electromechanical device, characterized in that, It includes the following steps: Stp1. Determine the expected life correlation attribute sequence according to the type of electromechanical equipment ; Based on the foregoing associated attribute sequence, from the electromechanical equipment to be analyzed the monitoring data is used to determine the existing monitoring data containing each attribute in the associated attribute sequence as the data to be analyzed; Select historical monitoring data covering the corresponding period or stage from the historical monitoring data of existing samples of the same type of electromechanical equipment as the initial reference data; ​ Stp2. Preferably, establish a database with the selected data from the existing samples to form a sample database; select an appropriate data period based on the data acquisition cycle characteristics of the samples to be analyzed and the same type of devices. , ; where refers to the sampling period of any attribute element , and is the minimum period to ensure the validity of the sample data of the attribute element . For the sample attribute sequence for each attribute element in it, use a sliding time window to intercept the corresponding periodic monitoring data. Based on the sliding time window the intercepted attribute element generate a periodic monitoring data curve for the monitoring data, and perform similarity analysis with the periodic monitoring data curve of the corresponding attribute element of the sample to be analyzed to determine each attribute element a time window with the largest similarity of the periodic monitoring data, extract the periodic monitoring data within this time window as valid data and retain it to obtain the sample the attribute element of the sequence of attribute values closest within the corresponding period , ; plot the parameter change curve within the time window to determine the sample the attribute element of the periodic change rate ; Among them ; Similarly, obtain the sample to be analyzed for the parameter periodic change rate ; Calculate the samples to be analyzed in historical data with each existing sample for the deviation of the periodic change rate ; where represents the sampling data of the th type of parameter of the sample in the th acquisition period; is the sampling period of the attribute element Sort the existing samples according to the principle of increasing deviation of the cycle change rate from small to large Perform sorting, preferably select several existing samples, form an effective data set for each sample from the corresponding effective data, and organize each sample to establish an effective reference database containing all samples; Stp3. Establish the expressions of prediction weights and predicted lifetimes based on the temporal similarity of sample data, including obtaining the sample matrix to be analyzed expressed by multi-dimensional vectors from the data to be analyzed and the sample matrix to be analyzed ; ； ； Based on the samples of the previous step of the attribute elements of the periodic change rate Determine the periodic change matrix of the sample to be analyzed according to the calculation method and the matrix of the sample to be analyzed , where: ； ； Since the changes in the equipment life cycle are often accompanied by continuous changes in the equipment performance indicators, the attribute data at each time node cannot well reflect the temporal change trend of the corresponding attribute characteristics. According to experience, when the life stage of a certain attribute data is closer to the current life stage of the product to be analyzed, the life expectancy of the equipment under the current data is also closer to the life expectancy of the product to be analyzed. To reflect this kind of influence, a stage matching index is established to characterize the matching degree of the life stage between the sample and the product to be analyzed , , the stage matching index characterizes the proximity of the life cycle node of this data to the current life cycle node of the equipment to be analyzed. The closer it is, the larger the stage matching index is. On the contrary, the smaller it is. The faster the equipment life cycle changes, the larger the magnitude of the stage matching index is. On the contrary, the smaller it is; Based on the similarity measurement principle, determine the periodic change matrix of the sample to be analyzed and the reference sample matrix The similarity can be expressed as ; For the similarity Perform dimensionless processing to obtain the influence degree of the reference sample considering the life cycle matching degree relative to the sample to be analyzed ; ; To determine the influence weights of multiple reference samples on the sample to be analyzed in life prediction, while ensuring that the weight value range is , given a positive real number , the weight expression form of the reference samples relative to the sample to be analyzed is obtained; after normalization, its weight expression is ; where is the total number of reference samples; Then the predicted life of the sample to be analyzed based on the reference sample can be expressed as: ； Among them, represents the time node the expected remaining life of the sample to be analyzed at the moment , indicates the time node the remaining life of the reference sample at the moment .

2. The method for predicting the service life of the electromechanical equipment according to claim 1, characterized in that, The data to be analyzed and the initial reference data include each associated attribute in the associated attribute sequence historical monitoring data, and the original monitoring data includes multiple groups of original monitoring data under different working conditions of the normal operation of the electromechanical equipment.

3. The method for predicting the service life of the electromechanical equipment according to claim 1, wherein In the step Stp3, since the acquisition periods of each attribute element are different and the total amount of attribute data is also different, in the actual implementation process, the matrix parameters are optimized by zero-value filling and normalization processing methods to facilitate the operation and processing.

4. The method for predicting the service life of the electromechanical equipment according to claim 1, characterized in that In the step Stp1, to ensure the validity of each attribute element in the associated attribute sequence and ensure that the attribute elements at the screening can characterize the expected life of the electromechanical equipment, each element in the associated attribute is determined based on the following method: A1. First, obtain the original monitoring data and perform necessary preprocessing, including deleting and revising incorrect data and normalizing the data. The original monitoring data includes multiple groups of original monitoring data under different working conditions of the electromechanical equipment during normal operation; After performing necessary preprocessing on the original data, segment it to obtain a training data set and a validation data set; A2. Establish a multi-layer directed feature extraction model composed of multiple restricted Boltzmann machine network models (RBMs), where the restricted Boltzmann machine network model includes a visible layer and a hidden layer . The visible layer serves as the input side to obtain sample data, and the hidden layer serves as the output side to extract the main features of the samples and output them in the form of a probability distribution; To optimize the data feature extraction effect, enhance the integrity of the data, and avoid falling into local optima, the global weight change amount and the local weight change are used to control the weight change of the hidden layer neurons. The algorithm is optimized by increasing the number of hidden layers and the number of neurons in the hidden layer. Specifically: The condition for increasing the number of neurons in the hidden layer can be expressed as: ; The condition for increasing the number of hidden layers can be expressed as: ; where ; ; After iterations, the weight vector of the neurons in the hidden layer ; ; Refers to the global weight change of the neuron for the th training sample ; is the total number of samples, At the time, the number of samples that increased the global weight change in the previous iteration; where is the maximum global weight change to ensure that each hidden layer neuron can correspond to a training sample after the number of hidden layer neurons increases threshold function, Can be expressed as: ； is the maximum number of iterations, represents the slope at the corresponding position of the variable threshold function curve, is the maximum value of the threshold function, is the minimum value of the threshold function, represents the energy function, is the preset control threshold; represents the number of hidden layers, represents the maximum number of the current hidden layer; During the training process, analyze the weight changes of the neurons in the hidden layer If the condition for increasing the number of neurons in the hidden layer is met, add a new neuron based on this neuron and reset the parameters of the two neurons to zero. At the same time, analyze whether the hidden layer meets the condition for increasing the number of hidden layers. If it does, copy this hidden layer and keep the original parameters unchanged; A3. Based on the foregoing steps, establish and optimize a multi-layer directed feature extraction model, input the training data into the model for training, use the validation data set to optimize the training of the model to obtain the final model, and extract the data change characteristics of the sample health state change process through the model. Screen through the correlation between these feature attributes and the correlation features of the electromechanical equipment health state to determine the preferred associated attribute sequence.

Images