A damper monitoring anomaly identification method based on a fractional order neural network

Abstract

The application relates to the technical field of damper monitoring and anomaly identification, and discloses a damper monitoring anomaly identification method based on a fractional order neural network, which comprises the following steps: using a distributed sensor network to monitor the running state of the damper in real time and store the original data set, and using a wavelet transform-based adaptive filtering algorithm to perform filtering processing, then combining a short-time Fourier transform and a Hilbert transform to extract characteristic information and perform nonlinear quantization and discretization mapping; labeling the discrete monitoring data to obtain training samples, training a fractional order neural network model, and using multiple optimization strategies to optimize the model in stages during the training process; and using the trained model to identify abnormal data from to-be-identified monitoring data. The application accurately extracts signal characteristics in combination with a wavelet transform, a short-time Fourier transform and a Hilbert transform, optimizes the model in stages in combination with multiple optimization strategies, and improves the accuracy of anomaly detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of damper monitoring and abnormality identification, and particularly relates to a damper monitoring abnormality identification method based on a fractional order neural network. BACKGROUND

[0002] With the complexity and automation level of modern industrial equipment improving, dampers, as key components, are widely used in mechanical, bridge, and building fields, and abnormal changes in the operating state of the damers can directly affect the stability and safety of the system. However, because the damper is in a complex working environment for a long time, the vibration, temperature, acceleration, and other sensor data often have characteristics such as high dimensionality, nonlinearity, and strong noise interference, making it difficult for traditional abnormality monitoring methods to accurately identify fault signals. At present, existing technologies mainly rely on fault diagnosis methods based on Fourier transform, time-frequency analysis, or integer order neural networks, but these methods often have problems such as low information utilization rate, insufficient sensitivity to abnormal data, and unstable training when dealing with high-dimensional, multi-modal sensor data. In particular, when facing long-tail distribution of abnormal data or complex environmental noise, the recognition accuracy and robustness of traditional methods are difficult to guarantee. In addition, most abnormality detection systems rely on offline analysis, making it difficult to achieve real-time monitoring and intelligent early warning of the damper state, thereby affecting the maintenance efficiency and fault prevention ability of the equipment. Therefore, there is an urgent need for an intelligent identification method that can efficiently process high-dimensional, complex sensor data and has strong abnormality detection ability and real-time monitoring ability to improve the monitoring accuracy of the operating state of the damper and enhance the safety and reliability of the system. SUMMARY

[0003] The purpose of the present application is to provide a damper monitoring abnormality identification method based on a fractional order neural network to solve the problems of poor adaptability, low abnormality detection accuracy, insufficient training stability, and limited real-time monitoring ability in the processing of high-dimensional, multi-sensor data in the prior art, realize real-time monitoring and abnormality early warning of the operating state of the damper, and improve the safety and operation and maintenance efficiency of the equipment.

[0004] To achieve the above-mentioned purpose, the present application provides the following solutions.

[0005] A damper monitoring abnormality identification method based on a fractional order neural network, the method comprising the following steps:

[0006] S1, data acquisition and storage: a distributed sensor network is used to monitor the operating state of the damper in real time, and each sensor monitoring data is stored in a structured text format to obtain an original data set;

[0007] S2, data preprocessing and discretization: for the original data set, an adaptive filtering algorithm based on wavelet transform is used for filtering processing, for the filtered signal, the characteristic information is extracted by joint short-time Fourier transform and Hilbert transform, and nonlinear quantization and discretization mapping are carried out, and discrete monitoring data is obtained;

[0008] S3, fractional order neural network model training: the discrete monitoring data is labeled to obtain training samples, and the training samples are used to train the fractional order neural network model, and a plurality of optimization strategies are used to optimize the model in stages during the training process, the optimization strategies include data distribution weighting adjustment, gradient accumulation mechanism and dynamic attenuation strategy, fractional calculus theory, parameter space rotation method, gradient normalization and sparsity adjustment strategy;

[0009] S4, damper monitoring abnormality identification: the trained fractional order neural network model is used to identify abnormal data for the monitoring data to be identified.

[0010] Further, in the S1, the distributed sensor network includes a vibration sensor, a temperature sensor and an accelerometer arranged inside the damper;

[0011] The data acquisition mode of the distributed sensor network is continuous acquisition at a fixed sampling rate, and the data is uploaded synchronously through an industrial bus to ensure accurate correspondence between the data timestamp and the sensor information;

[0012] The structured text format includes JSON and XML formats, and the stored content includes the original values of each sensor, the acquisition time and the sensor number information, which are used as the original data set for subsequent data processing.

[0013] Further, in the S2, data preprocessing and discretization, specifically comprising:

[0014] S201, for the original data set, an adaptive filtering algorithm based on wavelet transform is used for filtering processing, specifically: the original signal is decomposed by multi-scale, and the low noise signal is reconstructed by using the minimum reconstruction error criterion, so that the main component of the original signal is obtained;

[0015] S202, the filtered signal is processed by joint short-time Fourier transform and Hilbert transform, and the time domain amplitude, frequency energy distribution and phase characteristics of the signal are extracted;

[0016] S203, based on the extracted continuous characteristics, a segmented nonlinear mapping algorithm is used for nonlinear quantization and discretization mapping, each characteristic value is mapped to a preset discrete set, and discrete monitoring data is obtained.

[0017] Further, in the S3, the discrete monitoring data is labeled to obtain training samples, specifically comprising:

[0018] The discrete monitoring data is labeled as normal data, slightly abnormal data and serious abnormal data.

[0019] Further, in the S3, a plurality of optimization strategies are adopted in the training process to optimize the model in stages, specifically including:

[0020] S301, in the initialization stage, a standard neural fractional order neural network initialization method is adopted and the initial weight is adjusted through data distribution weighting, and the dynamic adjustment of the weighting factor is realized according to the two norm of the initial feature vector;

[0021] S302, in the gradient accumulation and attenuation stage, the attenuation factor of each iteration training and the accumulated nonlinear adjustment gradient of each iteration are determined by adopting the gradient accumulation mechanism and the dynamic attenuation strategy, the weight is updated according to the accumulated gradient, and the updated weight matrix of each iteration is obtained;

[0022] S303, in the fractional order gradient update stage, for the detection of subtle changes of abnormal data in the damper monitoring abnormal data identification task, the gradient update is corrected by adopting the fractional order calculus theory in the calculation of the weight change, and the efficient capture of complex nonlinear relationship and long tail characteristics is realized;

[0023] S304, in the parameter space rotation strategy stage, by adopting the parameter space rotation method, the parameter matrix is rotated and transformed before each weight update, the multi-directional search is realized, and the exploration of the parameter space in the damper monitoring abnormal data identification task is optimized;

[0024] S305, in the high-dimensional feature adjustment and gradient constraint stage, based on the gradient normalization and sparsity adjustment technology, the loss function gradient of each iteration is first determined, the feature vector is updated according to the gradient sparsity, and the feature contribution is dynamically optimized in each iteration training process to obtain the updated feature vector;

[0025] S306, in the dynamic learning rate adjustment stage, by adopting the nonlinear adjustment mechanism based on the training process, loss change and accumulated gradient information, the learning rate of the fractional order neural network is dynamically attenuated;

[0026] S307, in the iterative training process of the fractional order neural network model, based on the current weight matrix and feature vector, the forward propagation calculation is performed to generate the prediction output, and according to the difference between the prediction result and the real label, the loss function of the iteration is calculated;

[0027] S308, repeat the above steps S302-S307 until the preset stop iteration condition is met, which means that the model training is completed.

[0028] Furthermore, in S302, the calculation method for the decay factor during each training iteration is expressed as follows:

[0029]

[0030] In the formula, For the first The decay factor during the next iteration of training; The adjustment factor characterizes the strength of the effect of the current loss gradient on the decay. For the first The gradient of the loss function in each iteration represents the change in training error;

[0031] The calculation method for the cumulative nonlinear adjustment gradient in each iteration is expressed as follows:

[0032]

[0033] In the formula, For the first The nonlinear adjustment gradient accumulated over multiple iterations; For the first The gradient of the loss function in the next iteration. Indices are positive integers; It is a fractional exponent, representing the power exponent of nonlinear regulation; For the first The squared L2 norm of the gradient of the loss function in the next iteration represents the magnitude of the gradient.

[0034] The weight matrix after each iteration is calculated as follows:

[0035]

[0036] In the formula, For the first The weight matrix updated in the next iteration; For the first The weight change in the next iteration; For the first The weight matrix for the next iteration; For the first The rotation matrix during the next iteration of training represents the rotation transformation of the parameter space; The learning rate is the fractional-order neural network, representing the step size for each update. The cumulative nonlinear adjustment gradient.

[0037] Furthermore, in step S303, the formula for calculating the weight change is as follows:

[0038]

[0039] In the formula, the weight change amount of the first iteration; the weight change amount of the first iteration; the fractional order index; the weight change amount of the first iteration; the loss function gradient of the first iteration; the weight change amount of the first iteration; the dynamic adjustment factor of the first iteration, representing the adaptive adjustment of the gradient dependence relationship between different layers; represents the fractional order root operation on the cumulative result; represents the index of the number of iterations, the current number of iterations is the first the next number of iterations is the first

[0040] wherein, the dynamic adjustment factor of the first iteration is calculated as follows:

[0041]

[0042] wherein, is the dynamic adjustment factor of the first iteration; is the dynamic adjustment factor of the first iteration; is the L2 norm of the loss function gradient of the first iteration; is the L2 norm of the loss function gradient of the first iteration; is the adjustment factor, representing the control strength of the gradient influence range of different layers.

[0043] Further, the S304, before each weight update, performs a rotation transformation on the parameter matrix, specifically including: calculating a rotation matrix:

[0044]

[0045] wherein, is the rotation matrix during the first iteration training; is the rotation matrix during the first iteration training; is a nonlinear transformation matrix, representing the adaptive adjustment of the parameter space topology structure; is an identity matrix, representing the standard benchmark of the parameter space; is a rotation factor, representing the contribution weight of gradient information in rotation; is the loss function gradient of the first iteration, representing the current error change; is the loss function gradient of the first iteration, representing the current error change; is the transpose of the loss function gradient of the first iteration, representing the row vector form of the gradient; is the transpose of the loss function gradient of the first iteration, representing the row vector form of the gradient;

[0046] According to the parameter space topology structure and the gradient information, the rotation matrix is supplemented and adjusted by adopting a nonlinear transformation strategy, realizing the dynamic optimization of the parameter space structure, and the calculation mode is represented as: ​​

[0047]

[0048] wherein, is a nonlinear transformation matrix; is an exponential function, representing the nonlinear attenuation effect; is a hyper-parameter, representing the influence of gradient amplitude on the adjustment of the transformation matrix; is a contribution degree hyper-parameter, representing the degree of contribution of the gradient to the nonlinear transformation.

[0049] Further, in the S305, the calculation method of the loss function gradient of each iteration is represented as:

[0050]

[0051] wherein, is the loss function gradient of the th iteration; is a parameter update operation; is the L2 norm of the loss function gradient of the th iteration; is a small constant to prevent division by zero errors; is the zero norm of the loss function gradient of the th iteration, representing the number of non-zero gradient elements; is a sparsity adjustment factor, representing the amplification or inhibition degree of the influence on gradient sparsity;

[0052] The updated feature vector after iteration is calculated as:

[0053]

[0054] wherein, is the updated feature vector after the th iteration; is the feature vector of the th iteration, representing the previous feature state; is a feature adjustment factor, representing the influence degree of the gradient on the feature update; is a small constant to prevent division by zero errors; is a feature update adjustment factor, representing the contribution of gradient sparsity to feature adjustment.

[0055] Further, in the S308, the preset stopping iteration condition includes:

[0056] 1) The loss change rate of the th iteration is less than the preset threshold , represented as:

[0057] ;

[0058] wherein, is the loss function of the first iteration, is the loss function of the first iteration;

[0059] 2) the maximum number of iterations is reached ;

[0060] 3) the L2 norm of the loss function gradient is continuously below a threshold , indicating that the parameter space is close to a stationary state.

[0061] According to the specific embodiments provided by the application, the damper monitoring anomaly identification method based on the fractional order neural network provided by the application discloses the following technical effects:

[0062] By combining wavelet transform, short-time Fourier transform and Hilbert transform, multi-modal feature extraction of the signal is realized, and the accuracy of signal feature extraction is improved; data distribution weighted initialization is adopted, so that the initial weight of the model can adapt to the amplitude difference of different sensor data, improve the model stability and training convergence speed, and through gradient accumulation, dynamic attenuation and parameter space rotation optimization strategies, the robustness of the model to noise data is enhanced, and the local optimal trap is avoided. At the same time, the application constructs an adaptive learning rate adjustment mechanism based on nonlinear regulation, so that the model can dynamically optimize the step size in the training process, improve the convergence speed and abnormal identification accuracy, and realize real-time monitoring and abnormal early warning of the damper operating state by means of a distributed sensor network, thereby improving the safety and operation efficiency of the equipment. BRIEF DESCRIPTION OF DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0064] Figure 1 is the flow chart of the damper monitoring anomaly identification method based on the fractional order neural network of the present application;

[0065] Figure 2 is the training process comparison chart of different initialization methods of the embodiments of the present application;

[0066] Figure 3 is the detection performance comparison chart of different noise levels of the embodiments of the present application;

[0067] Figure 4Figures of the capturing ability of long tail features of the embodiment of the present application, wherein (a) is a figure of the capturing ability of long tail distribution features in damper monitoring data of the fractional gradient update mechanism proposed by the present application, (b) is a column chart for performance comparison between the traditional neural network method and the method of the present application, and (c) is a scatter plot for comparison of detection results between the traditional neural network method and the method of the present application.

[0068] Figure 5 Figures of model robustness analysis of the traditional neural network method and the method of the present application under high-dimensional feature space. DETAILED DESCRIPTION

[0069] Embodiments of the present application are described in detail below, which are intended to explain the present application and cannot be understood as a limitation of the present application. If specific techniques or conditions are not mentioned in the embodiments, the techniques or conditions described in the literature in the art or according to the product instructions are used. If the manufacturer of the materials or instruments is not mentioned, it is a conventional product that can be obtained on the market.

[0070] As shown in Figure 1 The damper monitoring anomaly recognition method based on the fractional order neural network provided by the embodiment of the present application comprises the following steps:

[0071] S1, data acquisition and storage:

[0072] The distributed sensor network is used to monitor the running state of the damper in real time, and the monitoring data of each sensor is stored in a structured text format to obtain an original data set.

[0073] Specifically, the distributed sensor network is used to monitor the running state of the damper in real time, and the data acquisition is derived from the monitoring data of multiple sensors such as vibration sensors, temperature sensors and accelerometers inside the damper.

[0074] The data acquisition mode is continuous acquisition with a fixed sampling rate, and the synchronous uploading of data is realized through an industrial bus to ensure the accurate correspondence of the data timestamp and the information of each sensor.

[0075] The collected data are stored in a structured text format, preferably in JSON or XML format, and the storage content includes the original values of each sensor, the acquisition time and the sensor number and other information, which serves as the original basis for subsequent data processing.

[0076] S2, data preprocessing and discretization:

[0077] For the original data set, an adaptive filtering algorithm based on wavelet transform is used for filtering processing, the characteristic information of the filtered signal is extracted by combining short-time Fourier transform and Hilbert transform, and nonlinear quantization and discretization mapping are performed to obtain discrete monitoring data.

[0078] For the high-frequency noise generated by the damper in the working process under the interference of working conditions, an adaptive filtering algorithm based on wavelet transform is adopted, which is specifically realized as follows: the original signal is subjected to multi-scale decomposition, the low-noise signal is reconstructed by using the minimum reconstruction error criterion, so that the main component of the signal is obtained.

[0079] Further, the filtered signal is subjected to joint processing of short-time Fourier transform (STFT) and Hilbert transform, and the time-domain amplitude, frequency-domain energy distribution and phase information of the signal are extracted.

[0080] Further, based on the above-mentioned extracted continuous features, nonlinear quantization and discretization mapping are performed, and the present application adopts a segmented nonlinear mapping algorithm to map each feature value into a preset discrete set, which is represented as:

[0081]

[0082] In the formula, represents the discretized feature value; is the original feature to be discretized; is the value after the first discrete mapping; is the value after the second discrete mapping; is the value after the third discrete mapping; is the first threshold parameter adaptively determined based on data distribution; is the second threshold parameter adaptively determined based on data distribution.

[0083] Through the above steps, the obtained discrete monitoring data, in an embodiment, includes: represents the amplitude measurement value, represents the frequency-domain energy measurement value, represents the phase measurement value, represents the time-domain statistical measurement value, represents the speed measurement value, represents the power measurement value, represents the quantization error measurement value, represents the modal parameter measurement value, represents the noise level measurement value, represents the external interference measurement value.

[0084] It should be noted that the present embodiment is only for illustrating one kind of data format and kind, and in actual application, the number of data attributes is usually more than 10, and the number of data attributes may reach dozens or even hundreds.

[0085] S3, fractional order neural network model training:

[0086] The discrete monitoring data is labeled to obtain training samples, and the fractional order neural network model is trained using the training samples. Various optimization strategies are used to optimize the model in stages during the training process. The optimization strategies include data distribution weighting adjustment, gradient accumulation mechanism and dynamic attenuation strategy, fractional calculus theory, parameter space rotation method, gradient normalization and sparsity adjustment strategy.

[0087] Specifically, the discrete monitoring data is labeled, and artificial labeling can be used. In embodiments, the labeled categories include:

[0088] Class A: normal data

[0089] Class B: slightly abnormal data

[0090] Class C: severe abnormal data

[0091] Each data sample is finely labeled according to the above attributes and labeled categories to ensure that the model can fully distinguish the subtle changes in the working state of the damper during training.

[0092] In the damper monitoring abnormal data identification task, a fractional order neural network model based on a gradient accumulation attenuation factor is used, and the model training process is as follows:

[0093] S301, in the initialization stage, the standard neural fractional order neural network initialization method is used and the initial weight is adjusted by data distribution weighting to ensure that the initial weight adapts to the discontinuity of fractional derivative and the complexity of high-dimensional data, and at the same time balances the influence of each feature in the damper monitoring abnormal data identification task. The calculation method is represented as:

[0094]

[0095] In the formula, is the weight matrix of the fractional order neural network at initialization; is the number of input nodes of the current layer (a positive integer), representing the dimension of the input data; is a random variable sampled from a standard normal distribution, representing random initialization; is a data feature weighting factor, representing the importance of each feature in the initial stage.

[0096] Further, since the damper monitoring data is derived from multiple sensors, the feature dimension is high and there may be amplitude differences. In order to take into account the importance of vibration, temperature, acceleration and other attributes to abnormal identification, a weighted random initialization method is used during initialization. By calculating the weighting factor of the data feature according to the overall amplitude of the initial feature vector (using the two norm of the vector), the initial weight of the model reflects the dispersion of each sensor data and balances the influence of each feature. The two norm of the initial feature vector is used to realize dynamic adjustment of the weighting factor, and the calculation of the data feature weighting factor is represented as:

[0097]

[0098] In the formula, is the data feature weighting factor; is the L2 norm of the initial feature vector characterizing the overall amplitude of the feature; is a small constant to prevent division by zero error, preferably set to 0.0001.

[0099] Unlike traditional random initialization, the score order derivative is more robust when dealing with high-dimensional and discontinuous data through data distribution weighting adjustment, so that the model is sensitive to small changes in the damper state at the initial stage. For example, vibration data may have large amplitude and rapid changes, while temperature data is relatively smooth. The weighting factor can make both types of data reasonably expressed in the same network.

[0100] To verify the advantages of the initialization method of the application, the following experiments are performed: the fractional stage weighting initialization method in the application is compared with Xavier initialization and He initialization.

[0101] The experiment aims to verify the superiority of the data distribution weighted initialization strategy adopted by the fractional order neural network compared with the traditional initialization method. By comparing the performance of the standard initialization method and the weighted random initialization method proposed in the application during the training process, it is found that the fractional order neural network exhibits faster convergence speed at the initial stage of training, and the final detection accuracy is significantly higher than that of the traditional method. As shown in Figure 2 the experimental curve shows that the traditional method still has obvious fluctuations in the middle of the training, while the learning trajectory of the fractional order neural network is more stable, verifying the effectiveness of dynamically adjusting the weighting factor through the norm of the feature vector. The initialization strategy of the application can automatically balance the influence of multi-dimensional features such as vibration and temperature according to the amplitude difference of sensor data, so that the model has the ability to capture subtle changes of abnormal signals from the initial stage of training, and at the same time, it relieves the discontinuity challenge of fractional order derivative when dealing with high-dimensional data.

[0102] S302, in the gradient accumulation and attenuation phase, by adopting the gradient accumulation mechanism and the dynamic attenuation strategy, the gradient information of each iteration training is gradually accumulated, so as to effectively suppress the influence of noise and abnormal samples in the damper monitoring abnormal data identification task, and ensure that the gradient update is neither too excited nor loses effective information, and the calculation method is represented as:

[0103]

[0104] In the formula, is the attenuation factor in the first iteration training; is an adjustment factor, representing the influence intensity of the current loss gradient on attenuation, preferably set to 0.3; is the loss function gradient of the first iteration, representing the training error change.

[0105] Further, according to the nonlinear adjustment mode, the gradient is accumulated to alleviate the gradient explosion problem. In the update, all iterations of the gradient are accumulated by nonlinear, which can alleviate the gradient explosion problem. Especially for the long tail abnormality that may exist in the sensor data, it can help the network to better capture the subtle changes, so as to identify the real abnormal data in the noise background. The calculation method is represented as:

[0106]

[0107] In the formula, is the nonlinear adjustment gradient accumulated in the first iteration; is the loss function gradient of the first iteration (n is a positive integer index); is a fractional order index, representing the power index of nonlinear adjustment, preferably set to ; is the L2 norm square of the loss function gradient of the first iteration;

[0108] Further, the weight is updated according to the accumulated gradient, so that the weight can stably reflect the historical gradient information. In the damper monitoring, the collected data often contains noise and abnormal samples. By adopting the gradient accumulation mechanism and the dynamic attenuation strategy, the system can gradually accumulate the gradient information in each iteration, and adjust the attenuation factor according to the size of the current loss gradient, so as to suppress the excessive influence of abnormal noise on weight update. The calculation method is represented as:

[0109]

[0110] In the formula, is the weight of the first​ the updated weight matrix after the second iteration; the weight change amount for the first iteration; the weight matrix for the first iteration; the rotation matrix during the first iteration training, representing the rotation transformation of the parameter space; the learning rate of the fractional order neural network, representing the step size of each update, preferably set to 0.02; the accumulated nonlinear adjustment gradient.

[0111] To verify the anti-noise ability of gradient accumulation and attenuation phase, the experimental analysis is as follows:

[0112] This experiment focuses on the abnormal detection robustness of fractional order neural network in noisy environment, and compares it with various classical machine learning methods (including LSTM network, SVM, random forest). Through simulating the gradually increasing noise interference in sensor data, it is observed that the detection performance of traditional methods decreases sharply when the noise reaches a certain level, while the fractional order neural network still maintains stable recognition ability. As shown in Figure 3 , it shows the inhibition effect of gradient accumulation and attenuation mechanism on abnormal samples, and the nonlinear capture characteristics of fractional calculus on long-tail features. Especially in high noise scene, traditional methods are easily disturbed by redundant features, leading to misjudgment, while the method of the present application can strengthen the response to key signals such as sudden vibration, while suppressing the interference of irrelevant factors such as temperature fluctuation. The experimental results directly reflect the advantages of gradient normalization and sparsity adjustment technology in high-dimensional data processing, proving that this method is more suitable for complex and variable damper monitoring scenes in practical engineering.

[0113] S303, in the fractional gradient update phase, for the detection of subtle changes in abnormal data in damper monitoring abnormal data identification task, in the weight change amount calculation, by using fractional calculus theory to correct the gradient update, realize the efficient capture of complex nonlinear relationship and long tail feature, the calculation mode is expressed as:

[0114]

[0115] In the formula, the weight change amount for the first iteration; the fractional order exponent; the loss function gradient for the first iteration; the dynamic adjustment factor for the first iteration, representing the adaptive adjustment of the gradient dependence relationship between different layers; represents a fractional order root operation on the cumulative result; represents an index of the number of iterations, the current number of iterations is the th, and the next number of iterations is the th.

[0116] Further, according to the gradient amplitude of each layer, the calculation of the dynamic adjustment factor is realized, which ensures the adaptive response to abnormal data in the weight update process of each layer. The dynamic adjustment factor adaptively adjusts the gradient dependence relationship between different layers, thereby ensuring the sensitive capture of subtle changes in abnormal data. The calculation method is represented as:

[0117]

[0118] In the formula, is the dynamic adjustment factor of the th iteration; is the L2 norm of the loss function gradient of the th iteration; is an adjustment factor, representing the control strength of the gradient influence range of different layers, and is preferably set to 3.

[0119] S304, in the parameter space rotation strategy stage, by adopting the parameter space rotation method, the parameter matrix is rotated and transformed before each weight update, realizing multi-directional search, thereby optimizing the exploration of the parameter space in the damper monitoring abnormal data identification task, avoiding local optimal phenomenon, and improving the model robustness and convergence. The calculation method is represented as:

[0120]

[0121] In the formula, is the rotation matrix in the th iteration training; is a nonlinear transformation matrix, representing the adaptive adjustment of the parameter space topology structure; is an identity matrix, representing the standard benchmark of the parameter space; is a rotation factor, representing the contribution weight of gradient information in rotation, and is preferably set to 0.5; is the loss function gradient of the th iteration, representing the current error change; is the transpose of the loss function gradient of the th iteration, representing the row vector form of the gradient.

[0122] Further, according to the parameter space topology structure and gradient information, the rotation matrix is supplemented and adjusted by adopting the nonlinear transformation strategy, realizing the dynamic optimization of the parameter space structure, and the calculation method is represented as:

[0123]

[0124] wherein, is a nonlinear transformation matrix; is an exponential function, representing the nonlinear attenuation effect; is a hyperparameter, representing the influence of gradient amplitude on the adjustment of the transformation matrix, preferably set to 0.5; is a contribution degree hyperparameter, representing the degree of contribution of the gradient to the nonlinear transformation, preferably set to 0.2.

[0125] S305, in the high-dimensional feature adjustment and gradient constraint phase, by adopting gradient normalization and sparsity adjustment technology, the gradient sparsity and inconsistency in high-dimensional data are solved, so that in the damper monitoring abnormal data identification task, the abnormal signal can be accurately identified and the interference of redundant features is suppressed, and the calculation mode is represented as:

[0126]

[0127] wherein, is the loss function gradient of the th iteration; is a parameter update operation; is the L2 norm of the loss function gradient of the th iteration; is a small constant to prevent division by zero error, preferably set to 0.0001; is the zero norm of the loss function gradient of the th iteration, representing the number of non-zero gradient elements, using the zero norm of the gradient to measure the number of non-zero elements, and dynamically adjusting the contribution of each dimension during feature update, so that the fractional order neural network can strengthen the response to sparse and key signals (such as sudden vibration anomalies) while suppressing irrelevant or redundant information interference when identifying anomalies; is a sparsity adjustment factor, representing the amplification or suppression intensity of the influence on gradient sparsity, preferably set to 0.2.

[0128] Further, according to the gradient sparsity adjustment feature vector update, the feature contribution is dynamically optimized in each iteration training process, the redundant features are suppressed and the abnormal data detection effect is strengthened, in the high-dimensional data collected by distributed sensors, some features may have redundancy or noise, by normalizing the gradient and introducing the sparsity adjustment strategy, the truly useful abnormal signals and redundant features can be accurately distinguished, and the calculation mode is represented as:

[0129]

[0130] wherein, is the loss function gradient of the the feature vector updated in the (n-1)-th iteration; the feature vector updated in the (n-1)-th iteration; the feature vector updated in the (n-1)-th iteration; is a feature adjustment factor, representing the influence degree of the gradient on the feature update, preferably set to 0.4; is a small constant to prevent division by zero errors, preferably set to 0.0001; is a feature update adjustment factor, representing the contribution of gradient sparsity to feature adjustment, preferably set to 0.4.

[0131] S306, in the dynamic learning rate adjustment stage, by adopting a nonlinear adjustment mechanism based on training progress, loss change and cumulative gradient information, the learning rate of the fractional order neural network is dynamically decayed, ensuring that the step size gradually decreases in the damper monitoring abnormal data identification task in the later training stage, avoiding gradient shock and explosion, and the calculation method is represented as:

[0132]

[0133] wherein, is the learning rate of the fractional order neural network in the n-th iteration training; is the initial learning rate of the fractional order neural network, representing the step size in the initial training, preferably set to 0.1; is the initial learning rate of the fractional order neural network, representing the step size in the initial training, preferably set to 0.1; is the initial learning rate of the fractional order neural network, representing the step size in the initial training, preferably set to 0.1; is the maximum training iteration (positive integer); is a dynamic adjustment factor based on error, representing the influence of the gradient L2 norm on the learning rate adjustment of the fractional order neural network, preferably set to 0.1.

[0134] S307, in the iterative training process of the fractional order neural network, based on the current weight matrix and the feature vector, forward propagation calculation is performed to generate the predicted output, and based on the difference between the predicted result and the real label, the loss function of the n-th iteration is calculated, represented as:

[0135]

[0136] wherein, is the loss function of the n-th iteration; is the loss function of the n-th iteration; is a specific loss calculation, preferably implemented by using cross-entropy loss calculation method to quantify the difference between the predicted result and the actual state, so that each iteration can continuously optimize the network based on the real abnormal signal; is the predicted value of the n-th iteration, is the predicted value of the n-th iteration, is the real label.

[0137] ​Further, the termination of the iteration training of the above steps S302-S307 needs to satisfy one of the following conditions:

[0138] 1) The loss change rate of the next iteration (such as ) is less than a preset threshold (such as ), which is expressed as:

[0139] ;

[0140] In the formula, is the loss function of the nth iteration.

[0141] 2) The maximum number of iterations is reached (such as ).

[0142] 3) The L2 norm of the loss function gradient is continuously lower than a threshold (such as ), indicating that the parameter space is close to a steady state.

[0143] S4, damper monitoring anomaly identification:

[0144] The trained fractional order neural network model is used to identify abnormal data from the monitoring data to be identified.

[0145] The specific process is as follows:

[0146] S401, through the data acquisition system established in the early stage, the newly generated sensor data is preprocessed and discretized to form a feature vector in the same format as the training data. These data are input into the trained neural network in real time.

[0147] S402, the model performs forward propagation on the input data, automatically calculates the transmission and nonlinear mapping between nodes of each layer, and generates a prediction result, which is one of the categories of "normal data", "slightly abnormal data" or "serious abnormal data".

[0148] S403, the identified abnormal data is fed back to the monitoring system for subsequent analysis and processing. For serious abnormal conditions, the system will trigger an alarm mechanism to take timely maintenance or adjustment measures to ensure the safe and stable operation of the damper.

[0149] Figure 4 ​​​​​Fig. 2 (a) is a schematic diagram of the capturing ability of fractional gradient update on long-tail characteristics, verifying the capturing ability of the fractional gradient update mechanism of the trained fractional order neural network model of the application on the long-tail distribution characteristics in the damper monitoring data. In the real damper monitoring scene, the abnormal data often presents long-tail distribution characteristics, and most of the data are concentrated in the normal range, and a small amount of key abnormal points are distributed in the edge area of the characteristic space. The experiment compares the performance of the traditional neural network method and the method described in the application in the long-tail characteristic detection task, and focuses on evaluating the abnormal detection ability of three key sensor characteristics (amplitude feature, phase feature, and energy feature). The experimental data configuration is shown in Table 1.

[0150]

[0151] In the characteristic space distribution diagram, the traditional method can only detect abnormal points with obvious amplitude characteristics, and the subtle abnormal response of phase characteristics and energy characteristics is insufficient, and the method of the application comprehensively covers the abnormal points of three types of characteristics of amplitude characteristics, phase characteristics and energy characteristics. For example, Figure 4 The performance comparison bar chart shown in Fig. 2 (b) shows that the method of the application is significantly better than the traditional method in precision, recall rate and F1 score, especially in the recall rate index, which directly reflects the capturing ability of the fractional gradient update mechanism on sparse abnormal characteristics. For example, Figure 4 The detection result scatter plot shown in Fig. 2 (c) further confirms that most of the data points are above the reference line, indicating that the abnormal score of the same sample is higher in the method of the application, and there is a dense distribution in the performance improvement area (green area), which makes the system can discover the potential failure of the damper earlier.

[0152] In order to evaluate the robustness of the method of the application in the high-dimensional feature space, the damper monitoring data is usually derived from a multi-sensor network, and the feature dimension can reach hundreds of dimensions. The traditional method faces the problem of dimension disaster when processing high-dimensional data. The application compares the performance change of the traditional neural network and the method of the application in the dimension growth process from 10-dimensional to 200-dimensional feature space, and analyzes the accuracy stability and anti-overfitting ability, as shown in Fig. 2 (d). Figure 5

[0153] Figure 5 ​The experimental results show the performance changes of the two methods with the increase of dimensions. The performance curve (blue) of the traditional method decreases rapidly with the increase of dimensions, while the performance curve (red) of the present method remains stable. The actual performance curve shows that when the feature dimension reaches 200 dimensions, the accuracy of the traditional method decreases significantly, while the present method only shows slight fluctuations. The performance stability index (Z axis) further shows that the fluctuation range of the present method in high-dimensional space is significantly smaller than that of the traditional method. Based on this, it can be shown that the gradient normalization technique effectively solves the gradient sparsity problem in high-dimensional data, ensuring that each dimension feature is updated evenly. In addition, the sparsity adjustment technique suppresses the interference of redundant features by dynamically adjusting the feature contribution. The projection line (blue dashed line) of the traditional method in the figure shows obvious deviation, while the projection line (red dashed line) of the present method remains stable, verifying the optimization effect of the parameter space rotation strategy on the exploration of high-dimensional features.

[0154] In summary, the damper monitoring anomaly identification method based on the fractional order neural network provided by the present application adopts a fractional order neural network with a gradient accumulation decay factor, combines fractional calculus during gradient update, improves the processing capability of high-dimensional data, and enhances the identification accuracy of abnormal data; by dynamically adjusting the factor, the gradient dependency relationship between different layers is adaptively adjusted, thereby enhancing the model's ability to capture subtle changes in abnormal data; a joint method of wavelet transform + short-time Fourier transform + Hilbert transform is used to extract amplitude, frequency energy distribution and phase information, improving the accuracy of signal feature extraction; the characteristics are discretized by the piecewise nonlinear mapping algorithm, so that the model can better learn the abnormal patterns under different data distributions; a weighted random initialization method is used, and the model initial weight can adapt to the amplitude difference of different sensor data according to the L2 norm of the data characteristics, improving the model stability and training convergence speed; the gradient accumulation and nonlinear adjustment strategy accumulates the gradient information of multiple iterations, so that the model can more stably cope with noise data and long-tail anomalies, improving the accuracy of anomaly detection; high-dimensional feature adjustment and gradient constraint, through gradient normalization and sparsity adjustment technology, solve the gradient sparsity problem in high-dimensional data, so that the model can focus on key features more easily and reduce irrelevant information interference; a parameter space rotation method is used to rotate and transform before each weight update, realize multi-directional search, optimize the convergence of the model in high-dimensional space, and avoid falling into a local optimal solution; combined with error change and accumulated gradient information, the learning rate is dynamically adjusted, and the step size is gradually reduced in the later training stage, avoiding gradient shock or explosion, improving the training stability and final recognition effect.

[0155] In addition, the present application also provides a damper monitoring anomaly identification system based on a fractional order neural network, which is applied to execute the damper monitoring anomaly identification method based on the fractional order neural network described above, and comprises:

[0156] A data acquisition and storage module is configured to monitor the running state of the damper in real time by using a distributed sensor network, and store the monitoring data of each sensor in a structured text format to obtain an original data set;

[0157] A data preprocessing and discretization module is configured to filter the original data set by using an adaptive filtering algorithm based on wavelet transform, extract feature information from the filtered signal by using a short-time Fourier transform and a Hilbert transform, and perform nonlinear quantization and discretization mapping to obtain discrete monitoring data.

[0158] A fractional-order neural network model training module is configured to label the discrete monitoring data to obtain training samples, and train the fractional-order neural network model by using the training samples.

[0159] A damper monitoring anomaly identification module is configured to identify abnormal data from to-be-identified monitoring data by using the trained fractional-order neural network model.

[0160] The application further provides a computer readable storage medium having a computer program stored thereon, and the computer program is executed to implement the damper monitoring anomaly identification method based on the fractional-order neural network.

[0161] The remaining matters of the application are known technologies.

[0162] The above has described the embodiments of the application, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those skilled in the art without departing from the scope and spirit of the described embodiments. The selection of the terms used herein is intended to best explain the principles, practical application, or improvement of the technology in the market of the embodiments, or to enable other ordinary skilled persons in the art to understand the disclosed embodiments.

Claims

1. A damper monitoring anomaly identification method based on a fractional order neural network, characterized in that, The method comprises the following steps: S1, data acquisition and storage: a distributed sensor network is used to monitor the running state of the damper in real time, and each sensor monitoring data is stored in a structured text format to obtain an original data set; S2, data preprocessing and discretization: for the original data set, an adaptive filtering algorithm based on wavelet transform is used for filtering processing, feature information is extracted from the filtered signal by combining short-time Fourier transform and Hilbert transform, and nonlinear quantization and discretization mapping are performed to obtain discrete monitoring data; S3, fractional order neural network model training: the discrete monitoring data are labeled to obtain training samples, and the training samples are used to train the fractional order neural network model, and a plurality of optimization strategies are used to optimize the model in stages during the training process, the optimization strategies including data distribution weighting adjustment, gradient accumulation mechanism and dynamic attenuation strategy, fractional calculus theory, parameter space rotation method, gradient normalization and sparsity adjustment strategy; S4, damper monitoring anomaly identification: the trained fractional order neural network model is used to identify abnormal data from the monitoring data to be identified; In S3, the model is optimized in stages during the training process, specifically including: S301, in the initialization stage, the standard neural fractional order neural network initialization method is used to adjust the initial weight through data distribution weighting, and the dynamic adjustment of the weighting factor is realized according to the two-norm of the initial feature vector; S302, in the gradient accumulation and attenuation stage, the gradient accumulation mechanism and dynamic attenuation strategy are used to determine the attenuation factor at each iteration training and the nonlinear adjustment gradient accumulated at each iteration, the weight is updated according to the accumulated gradient to obtain the weight matrix updated at each iteration; S303, in the fractional order gradient update stage, for the detection of subtle changes in abnormal data in the damper monitoring abnormal data identification task, the gradient update is corrected by using the fractional calculus theory when calculating the weight change, and the complex nonlinear relationship and long tail feature are efficiently captured; S304, in the parameter space rotation strategy stage, the parameter space rotation method is used to rotate and transform the parameter matrix before each weight update to realize multi-directional search and optimize the exploration of the parameter space in the damper monitoring abnormal data identification task; S305, in the high-dimensional feature adjustment and gradient constraint stage, based on the gradient normalization and sparsity adjustment technology, the loss function gradient of each iteration is first determined, the feature vector is updated according to the gradient sparsity, and the feature contribution is dynamically optimized during each iteration training process to obtain the updated feature vector; S306, in the dynamic learning rate adjustment stage, a nonlinear adjustment mechanism based on training progress, loss change and accumulated gradient information is used to realize the dynamic attenuation of the learning rate of the fractional order neural network; S307, in the iteration training process of the fractional order neural network model, based on the current weight matrix and feature vector, forward propagation calculation is performed to generate a prediction output, and the difference between the prediction result and the true label is calculated to obtain the loss function of the iteration; S308, repeating the iteration of the above steps S302-S307 until a preset stop iteration condition is met, indicating that the model training is completed.

2. The method of claim 1, wherein, In S1, the distributed sensor network includes a vibration sensor, a temperature sensor, and an accelerometer arranged inside the damper. The data collection mode of the distributed sensor network is continuous collection at a fixed sampling rate, and the data is synchronously uploaded through an industrial bus to ensure accurate correspondence between the data timestamp and the sensor information. The structured text format includes JSON and XML formats, and the stored content includes the original values of each sensor, the collection time, and the sensor number information, serving as the original data set for subsequent data processing.

3. The method of claim 1, wherein, In S2, data preprocessing and discretization, specifically including: S201, for the original data set, an adaptive filtering algorithm based on wavelet transform is used for filtering processing, which is specifically implemented as: multi-scale decomposition is performed on the original signal, and a low-noise signal is reconstructed by using the minimum reconstruction error criterion, so as to obtain the main components of the original signal; S202, the filtered signal is subjected to joint processing of short-time Fourier transform and Hilbert transform, and the time domain amplitude, frequency domain energy distribution and phase characteristics of the signal are extracted; S203, based on the extracted continuous characteristics, a segmented nonlinear mapping algorithm is used for nonlinear quantization and discretization mapping, each characteristic value is mapped to a preset discrete set, and discrete monitoring data is obtained.

4. The method of claim 1, wherein, In S3, the discrete monitoring data is labeled to obtain training samples, specifically including: The discrete monitoring data is labeled as normal data, slight abnormal data and severe abnormal data.

5. The method of claim 1, wherein, In S302, the calculation method of the nonlinear adjustment gradient accumulated by each iteration is represented as: In the formula, is the nonlinearly adjusted gradient accumulated for the tth iteration; is the loss function gradient for the ith iteration, i is a positive integer index; a u is the fractional order index, representing the power index of the nonlinear adjustment; is the L2 norm square of the loss function gradient for the ith iteration, representing the gradient amplitude; The calculation method of the updated weight matrix of each iteration is represented as: In the formula, is the weight matrix updated for the tth iteration; is the attenuation factor at the tth iteration of training; is the weight change amount of the tth iteration; is the weight matrix of the (t-1)th iteration; is the rotation matrix at the tth iteration of training, representing the rotation transformation of the parameter space; η u is the learning rate of the fractional-order neural network, representing the step size of each update; is the accumulated nonlinearly adjusted gradient.

6. The method of claim 1, wherein, In S303, the weight change amount calculation formula is as follows: In the formula, is the weight change amount of the tth iteration; a u is the fractional order index; is the loss function gradient of the ith iteration; is the dynamic adjustment factor of the tth iteration, representing the adaptive adjustment of the gradient dependence relationship between different layers; represents the fractional order root operation on the cumulative result; t represents the index of the iteration number, the current iteration number is the t-1th, and the next iteration number is the tth.

7. The method of claim 1, wherein, In S304, the parameter matrix is subjected to rotation transformation before each weight update, specifically including: calculating a rotation matrix: In the formula, is the rotation matrix at the tth iteration training; T u is a nonlinear transformation matrix, representing the adaptive adjustment of the parameter space topology; I u is a unit matrix, representing the standard reference of the parameter space; α ecs is a rotation factor, representing the contribution weight of the gradient information in rotation; is the loss function gradient of the tth iteration, representing the current error change; is the transpose of the loss function gradient of the tth iteration, representing the row vector form of the gradient; According to the parameter space topology structure and gradient information, the rotation matrix is supplemented and adjusted by using a nonlinear transformation strategy to realize dynamic optimization of the parameter space structure.

8. The method of claim 1, wherein, In S305, the calculation method of the updated feature vector is represented as: In the formula, is the feature vector updated in the tth iteration; is the feature vector in the (t-1)th iteration, representing the last feature state; is the loss function gradient in the tth iteration; α vfr is a feature adjustment factor, representing the influence degree of the gradient on the feature update; ∈ u is a small constant to prevent division by zero error; β gr is a feature update adjustment factor, representing the contribution of the gradient sparsity to the feature adjustment.

9. The method of claim 1, wherein, In S308, the preset stop iteration condition includes: 1) the rate of change of loss for consecutive K iterations is less than a preset threshold ∈ L ; 2) a preset maximum number of iterations T is reached u ; 3) L2 norm of loss function gradient stays below threshold e G indicating that the parameter space is approaching a stationary state.

Images