Automatic dehumidification method in power equipment box

Abstract

The present invention relates to an automatic dehumidification method in a power equipment box, including collecting audio signals of environmental noise at different positions in the power equipment box and preprocessing; analyzing the acoustic characteristics of different materials in the power equipment box and establishing a propagation model; classifying and identifying environmental noise at different positions in different scenes; optimizing audio signals through adaptive noise reduction processing, extracting key volume characteristic parameters reflecting humidity status; estimating and obtaining real-time humidity values ​​according to key volume characteristic parameters; comparing the humidity value with a preset humidity threshold, and performing dehumidification processing if it exceeds the preset humidity threshold; during the dehumidification process, continuously monitoring humidity changes and detecting box abnormalities. The present invention realizes non-contact, real-time monitoring and intelligent control of humidity in the power equipment box through environmental noise analysis and acoustic characteristic modeling, and can accurately identify and precisely extract humidity-related volume characteristics, ensure the operation of dehumidification work, and improve the efficiency of dehumidification.

Description

Technical Field

[0001] The invention relates to the field of automatic control technology, and in particular to an automatic dehumidification method in a box of an electric power device. Background Art

[0002] The internal environment of the power equipment box is complex and changeable. The material differences in different locations and the interference of the external environment will affect the volume monitoring during the automatic dehumidification process. Humidity volume refers to the volume characteristics under different humidity. The characteristics of sound propagation under different humidity will change, causing the volume to change when the humidity changes. Humidity will affect the attenuation rate of sound.

[0003] In drier environments, sound decays faster because air molecules scatter and absorb sound waves more strongly, while in humid environments, water molecules in the air can absorb and reflect sound waves, slowing the decay of sound. This means that in environments with higher humidity, sound can travel farther, and humidity can also affect the clarity of sound. In environments with higher humidity, water molecules in the air interact with sound waves, causing sound to scatter and refract as it propagates, making the sound blurry or distorted. This effect is particularly noticeable in high-frequency sounds, which are more easily affected by air and water molecules.

[0004] Therefore, in an environment with high humidity, the clarity of sound may be reduced. Humidity also affects the frequency characteristics of sound. In an environment with high humidity, the frequency characteristics of sound may change due to the interaction between water molecules and sound waves. This change may be manifested as increased attenuation of high-frequency sounds, enhancement of low-frequency sounds, or changes in the balance of the entire spectrum, which may cause the sound to sound different from that in a dry environment. Differences in materials at different locations inside the cabinet will lead to different attenuation and reflection of sound propagation, resulting in deviations in the volume monitoring results at the same humidity. At the same time, noise interference in the environment outside the cabinet will also be mixed into the volume collection process, masking the characteristics of humidity volume. If the same noise reduction strategy is used in quiet scenes and noisy scenes, there will be problems such as excessive noise reduction leading to loss of humidity volume characteristics and insufficient noise reduction leading to obvious noise interference, which will ultimately affect the accurate judgment of humidity and the effect of automatic dehumidification. In addition, due to the limitations of the internal structure of the cabinet, the volume monitoring module cannot cover all areas, resulting in insufficient comprehensiveness and representativeness of the monitoring data. In summary, volume monitoring during automatic dehumidification in the power equipment box faces problems such as environmental noise interference, the influence of internal material differences, and incomplete monitoring coverage. Targeted optimization and improvement are needed to improve the dehumidification effect and the accuracy of humidity judgment. Summary of the invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide an automatic dehumidification method in the box of an electric equipment. The automatic dehumidification method in the box of an electric equipment realizes non-contact, real-time monitoring and intelligent control of the humidity in the box of the electric equipment through environmental noise analysis and acoustic characteristic modeling, and can accurately identify and precisely extract the volume characteristics related to humidity, ensure the operation of the dehumidification work, and improve the efficiency of dehumidification.

[0006] The automatic dehumidification method in a power equipment box of the present invention comprises:

[0007] S1, collecting audio signals of environmental noise at different positions in the box of the power equipment, and preprocessing the audio signals;

[0008] S2. Analyze the acoustic characteristics of different materials in the power equipment box and establish an acoustic propagation model;

[0009] S3. Classify and identify environmental noise at different locations and in different scenes in the power equipment box to obtain noise type and intensity characteristic information;

[0010] S4, performing adaptive noise reduction processing on the audio signal based on the environmental noise characteristics, the acoustic properties of the material and the attributes of different positions;

[0011] S5, extracting features from the audio signal after adaptive noise reduction to obtain key volume feature parameters reflecting humidity status;

[0012] S6. Estimate the humidity level in the power equipment box according to the key volume characteristic parameters to obtain real-time humidity values ​​of different materials at different locations;

[0013] S7, preset a humidity threshold, compare the estimated humidity value with the preset humidity threshold, and if the humidity exceeds the humidity threshold, perform dehumidification processing on the power equipment box;

[0014] S8. During the dehumidification process, continuously monitor the humidity changes at different locations in the power equipment box;

[0015] S9. During the dehumidification process, detect abnormal conditions of the power equipment cabinet.

[0016] Preferably, the audio signal collection in step S1 includes the following steps:

[0017] Environmental noise sensors are set at different positions in the power equipment box to collect the audio signal, and background noise characteristic parameters and humidity volume characteristics at various positions in the power equipment box are extracted through signal processing technology;

[0018] The preprocessing of the audio signal adopts an adaptive noise reduction algorithm;

[0019] The preprocessing of the audio signal includes removing noise interference and extracting effective humidity volume feature information.

[0020] Preferably, the step S2 further includes:

[0021] After establishing the acoustic propagation model, adjusting the parameters of the adaptive noise reduction algorithm by analyzing the attenuation and distortion degree of the audio signal caused by different materials;

[0022] The establishment of the acoustic propagation model comprises the following steps:

[0023] According to the internal structure layout and material information of the power equipment box, the finite element software is used to establish a three-dimensional acoustic scene model, divide the finite element grid, set boundary conditions and load conditions, input material property parameters, and calculate and simulate the acoustic characteristic parameters of different materials;

[0024] Extract key parameters of sound waves in each material, including propagation velocity, attenuation coefficient and reflection coefficient, and derive calculation results;

[0025] By analyzing the attenuation and distortion characteristic curves of different materials on audio signals, the influence of each material on the sound signal in different frequency ranges can be determined;

[0026] Constructing a correspondence between attribute parameters and their acoustic characteristic parameters, wherein the attribute parameters include a material acoustic characteristic database, recording material type, thickness, density and porosity;

[0027] The acoustic propagation model is established by combining the location of the sound source, the propagation path and the acoustic characteristics of the material.

[0028] Preferably, the step S3 specifically includes:

[0029] Environmental noise sensors are installed at different locations in the power equipment box to collect noise data samples under different working conditions and form a noise sample database;

[0030] The signal processing method is used to analyze the collected noise sample data in the time and frequency domain, extract the spectrum characteristics, energy characteristics and statistical characteristics of the noise signal, and construct a noise feature vector data set;

[0031] Using machine learning algorithms and combining them with noise feature vector datasets, we train machine learning models to classify and identify different types of noise, obtain noise type labels and intensity feature information, and form a noise identification knowledge base.

[0032] Preferably, the step S4 specifically includes:

[0033] In combination with the environmental noise characteristics, the acoustic properties of the material and the properties of different positions, the corresponding noise reduction algorithm and parameters are used to perform adaptive noise reduction processing on the audio signal;

[0034] The adaptive noise reduction processing of the audio signal includes eliminating noise interference and retaining humidity volume feature information.

[0035] Preferably, the step S5 specifically includes:

[0036] For the audio signal after adaptive noise reduction, a time-frequency analysis method is used to extract volume feature parameters reflecting the humidity state to form a high-dimensional humidity volume feature vector;

[0037] Perform multi-scale and multi-resolution analysis on the audio signal by wavelet packet decomposition, extract wavelet coefficient statistics at different frequency bands and time scales, and capture non-stationary humidity volume characteristics in the audio signal;

[0038] The speech feature extraction method is used to obtain the spectrum distribution characteristics reflecting the humidity state, and the dynamic characteristics of volume intensity changing with humidity are obtained;

[0039] The correlation and significance between each humidity volume feature and humidity state are evaluated through feature selection algorithm;

[0040] The maximum correlation minimum redundancy algorithm is used to first select the feature with the greatest correlation with the humidity state, and then select the feature with the least correlation with the selected feature from the remaining features until the preset number of features or correlation threshold is met;

[0041] The chi-square test was used to evaluate whether the feature was independent of the humidity state by calculating the difference between the actual observed frequency and the expected frequency of the feature value.

[0042] Preferably, the step S6 specifically includes:

[0043] Acquire the audio signals processed by adaptive noise reduction at different positions of different materials in the power equipment box, extract volume feature parameters reflecting the humidity state, construct a multi-dimensional humidity volume feature vector, and form an input data set for the humidity estimation model;

[0044] A long short-term memory neural network is used to build a deep learning humidity estimation model, and a Dropout layer is added between the LSTM layer and the fully connected layer to prevent the model from overfitting;

[0045] The mean square error is selected as the loss function of the humidity estimation model, and the Adam optimizer is used to adaptively adjust the model parameters. The model parameters with the lowest mean square error during the training process are saved as the optimal model.

[0046] The humidity volume feature data of different materials at different locations are divided into training set, validation set and test set by stratified sampling method. According to the distribution of humidity labels, the data set is divided into multiple humidity intervals, and then samples are randomly selected in each interval at a ratio of 8:1:1.

[0047] The random partitioning method is used to repeat the K-times data set partitioning, and the model is trained and evaluated each time. Finally, the average of the K-times results is taken as the final performance indicator of the model.

[0048] The training set is used to train the deep learning humidity estimation model, the validation set is used to evaluate the model performance and tune the parameters, and the test set is used to test the generalization ability of the model;

[0049] During the training process, the hyperparameters of the LSTM layer are dynamically adjusted to optimize the model's fitting ability and generalization performance. At the same time, an early stopping mechanism is introduced to terminate the training in advance to prevent the model from overfitting.

[0050] The trained deep learning humidity estimation model is deployed to an embedded device or a cloud server, and the audio signals of different materials at different positions in the power equipment box are received in real time. The humidity volume characteristic parameters are extracted and input into the model for humidity estimation to obtain the real-time humidity value.

[0051] Preferably, the step S7 specifically includes:

[0052] If the humidity estimation value of a certain location in the power equipment box output by the deep learning humidity estimation model is higher than the preset humidity threshold, a humidity abnormality alarm message is automatically generated, the time, location and humidity value of the abnormality are recorded, and the automatic dehumidification device at that location is triggered to start, and dehumidify the corresponding location in the power equipment box until the humidity returns to the theoretical normal humidity value.

[0053] Preferably, the step S8 further includes:

[0054] According to the humidity changes at different positions in the power equipment box, the working parameters of the automatic dehumidification device are adjusted in real time, and the working parameters include dehumidification intensity and duration.

[0055] Preferably, the step S9 further includes:

[0056] If no abnormality is detected, the original dehumidification working state is maintained;

[0057] If an abnormal situation is detected, the dehumidification work will be suspended, the alarm mechanism will be triggered, and an alarm message and data recording the abnormal situation will be sent.

[0058] The automatic dehumidification method in the box of an electric power equipment described in the present invention has the advantages that:

[0059] An automatic dehumidification method inside an electric equipment box of the present invention realizes non-contact, real-time monitoring and intelligent control of the humidity inside the electric equipment box through precise environmental noise analysis and acoustic characteristic modeling. It can accurately identify and reduce noise interference, and accurately extract humidity-related volume characteristics, and automatically dehumidify according to real-time changes in humidity; and through monitoring during the dehumidification process, it can ensure the operation of the dehumidification work, improve the efficiency of dehumidification, effectively extend the life of the electric equipment, and improve the operational stability and safety of the electric equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The present invention is a flowchart of an automatic dehumidification method in a power equipment box. DETAILED DESCRIPTION

[0061] like Figure 1 As shown, the automatic dehumidification method in the box of an electric power equipment described in the present invention comprises:

[0062] S1, collecting audio signals of environmental noise at different positions in the power equipment box, and preprocessing the audio signals;

[0063] S2. Analyze the acoustic characteristics of different materials in the power equipment box and establish an acoustic propagation model;

[0064] S3. Classify and identify environmental noise at different locations and in different scenes in the power equipment box to obtain noise type and intensity characteristic information;

[0065] S4, combining the environmental noise characteristics, the acoustic properties of the material and the properties of different positions, to perform adaptive noise reduction processing on the audio signal;

[0066] S5, extracting features from the audio signal after adaptive noise reduction to obtain key volume feature parameters reflecting the humidity state;

[0067] S6. Estimate the humidity level in the power equipment box based on the key volume characteristic parameters to obtain real-time humidity values ​​of different materials at different locations;

[0068] S7, preset a humidity threshold, compare the estimated humidity value with the preset humidity threshold, and if the humidity exceeds the humidity threshold, perform dehumidification processing on the power equipment box;

[0069] S8. During the dehumidification process, continuously monitor the humidity changes at different locations in the power equipment box;

[0070] S9. During the dehumidification process, detect abnormal conditions of the power equipment cabinet.

[0071] Furthermore, in this embodiment, the audio signal collection in step S1 includes the following steps:

[0072] Environmental noise sensors are set at different locations in the power equipment box to collect audio signals, and the background noise characteristic parameters and humidity volume characteristics at various locations in the power equipment box are extracted through signal processing technology;

[0073] The preprocessing of audio signals adopts an adaptive noise reduction algorithm; the adaptive noise reduction algorithm can select support vector machine, decision tree or naive Bayes algorithm;

[0074] The preprocessing of audio signals includes removing noise interference and extracting effective humidity volume feature information;

[0075] Specifically, according to the extracted background noise characteristic parameters and humidity volume characteristics, wavelet decomposition and Mel frequency cepstral coefficient (MFCC) method are used to extract the time-frequency domain characteristics of noise, and the optimal feature subset is selected through feature importance analysis and dimensionality reduction method;

[0076] Use machine learning models, such as support vector machines, decision trees, or naive Bayes, to classify and label the background noise characteristic parameters at each location, obtain sample data sets of different noise types and humidity volume characteristics, and use grid search and cross-validation techniques to optimize the hyperparameters of the machine learning model to improve the accuracy and robustness of noise classification;

[0077] Based on the labeled noise sample data set, the adaptive denoising model is trained. By dynamically adjusting the model hyperparameters and introducing the forgetting mechanism, the model can adapt to the changes in the noise environment, optimize the noise recognition and suppression capabilities of the denoising algorithm, and improve the adaptability and robustness of the denoising.

[0078] In the process of model updating, early stopping and regularization methods are used to prevent overfitting. The trained adaptive noise reduction model is deployed to the audio signal processing system of the power equipment box to pre-process the real-time collected audio signals and automatically identify and remove background noise interference.

[0079] The humidity volume feature information is extracted from the audio signal after adaptive noise reduction, and humidity anomaly analysis is further performed. The time-frequency analysis of the audio signal after noise reduction is performed to extract the key frequency band features reflecting humidity changes.

[0080] By analyzing the physical mechanism and mathematical model between electrical equipment noise and humidity changes, a mapping relationship model between humidity changes and key frequency band characteristics is established. This model is used to detect and warn of humidity anomalies in the key frequency band characteristics of real-time audio signals, and abnormal humidity problems in power equipment can be discovered.

[0081] During the noise reduction process, the noise reduction effect of the noise reduction algorithm is evaluated in real time, and the noise reduction effect is quantified using indicators, including signal-to-noise ratio (SNR), spectral entropy and speech quality index (PESQ). The evaluation indicator threshold is adaptively set according to the noise characteristics and noise reduction requirements in different environments;

[0082] When the noise reduction effect is lower than the preset threshold, the incremental learning and optimization of the noise reduction model is triggered. As the operating environment of the power equipment changes, the noise sample data at various locations is regularly collected, and an incremental learning strategy is adopted. The incremental learning strategy can choose online learning or small batch learning. The model is regularly updated with newly collected noise samples, and the adaptive noise reduction model is continuously trained and iteratively updated. The humidity anomaly detection and early warning algorithm is continuously improved to improve the accuracy and real-time performance of humidity anomaly analysis, and verify the feasibility and effectiveness of this method in actual power equipment operation and maintenance.

[0083] Here is an example:

[0084] Environmental noise sensors are set at different locations in the power equipment box. The environmental noise sensors can select vector sensors or MEMS microphone arrays to collect audio signals at a sampling rate of 48kHz and a quantization accuracy of 24 bits.

[0085] The collected audio signals are preprocessed, including DC removal, pre-emphasis and framing steps, and then the time-frequency domain features of each frame signal are extracted using wavelet decomposition and MFCC feature extraction algorithms;

[0086] Among them, the wavelet decomposition uses Daubechies wavelet function with 8 decomposition layers; MFCC feature extraction uses 40 triangular filters to extract 12-dimensional MFCC coefficients and 1-dimensional energy parameters, and through feature selection algorithms, such as the maximum relevance minimum redundancy (mRMR) algorithm based on mutual information, the 20-dimensional features that contribute most to noise classification and humidity volume feature information are selected as the input of the noise classifier;

[0087] The noise classifier adopts the support vector machine (SVM) model and uses the radial basis kernel function (RBF). Through grid search and 5-fold cross validation, the penalty factor C and kernel function parameter γ of SVM are optimized, and the classification accuracy on the noise sample data set reaches more than 95%;

[0088] According to the results of noise classification and humidity volume feature information, an adaptive noise reduction model is trained, such as a convolutional neural network (CNN) model. The CNN model contains 3 convolutional layers and 2 fully connected layers. The convolution kernel size is 3x3, the pooling layer size is 2x2, and the activation function uses ReLU.

[0089] During the training process, the Adam optimization algorithm was used, the initial learning rate was set to 0.001, and the learning rate was halved every 10 epochs. At the same time, L2 regularization and dropout techniques were used, the regularization coefficient was set to 0.01, and the dropout probability was set to 0.5 to prevent the model from overfitting;

[0090] After 50 epochs of training, the CNN model's noise reduction signal-to-noise ratio (SNR) on the validation set increased by more than 10dB. The trained CNN noise reduction model was deployed to the embedded audio processor in the power equipment box, and the real-time audio signal was processed with a frame length of 32ms to remove background noise interference;

[0091] Use the pre-trained humidity anomaly detection model to analyze the spectral characteristics of the denoised audio signal and extract the frequency band energy value related to humidity, such as the 2kHz-4kHz frequency band. Based on the mapping relationship between the frequency band energy value and humidity, determine whether the humidity value corresponding to the current audio signal exceeds the normal range.

[0092] When it is detected that the humidity value corresponding to 5 consecutive frames of audio signals exceeds the preset threshold (such as 80%), the humidity anomaly warning is triggered and the warning information is reported to the cloud management platform. At the same time, environmental noise samples and humidity anomaly samples are continuously collected, and the noise classifier, noise reduction model and humidity anomaly detection model are regularly incrementally trained and optimized.

[0093] Furthermore, in this embodiment, step S2 also includes:

[0094] After establishing the acoustic propagation model, the parameters of the adaptive noise reduction algorithm are adjusted by analyzing the attenuation and distortion of audio signals caused by different materials, so that it can adapt to the acoustic environment of different materials.

[0095] The establishment of the acoustic propagation model includes the following steps:

[0096] According to the internal structure layout and material information of the power equipment box, the finite element software is used to establish a three-dimensional acoustic scene model, divide the finite element grid, set boundary conditions and load conditions, input material property parameters, and calculate and simulate the acoustic characteristic parameters of different materials;

[0097] Extract the key parameters of sound waves in each material, including propagation speed, attenuation coefficient and reflection coefficient, and export the calculation results for subsequent analysis;

[0098] By analyzing the attenuation and distortion characteristic curves of different materials on audio signals, the influence of each material on the sound signal in different frequency ranges can be determined;

[0099] Construct the correspondence between attribute parameters and their acoustic characteristic parameters. The attribute parameters include a material acoustic characteristic database that records material type, thickness, density, and porosity for subsequent optimization of acoustic propagation model parameters and adaptive adjustment of noise reduction algorithms.

[0100] An acoustic propagation model is established by combining the location of the sound source, the propagation path, and the acoustic characteristics of the material;

[0101] Furthermore, the ray acoustics method is used to numerically simulate the sound propagation process in the power equipment box, and appropriate parameters such as ray density, reflection number and receiving radius are set. In combination with the influence of the sound absorbing material and sound insulation structure inside the box on sound propagation, a sound absorption and transmission model is established;

[0102] Compare the model calculation results with the measured data to verify the accuracy and effectiveness of the model. Combine the material acoustic feature database and the calculation results of the acoustic propagation model to dynamically adjust the filter parameters, gain coefficients and frequency compensation curves in the noise reduction algorithm.

[0103] According to the characteristics of the acoustic environment under different material combinations, an adaptive finite impulse response filter is used to optimize the filter order and cutoff frequency according to the acoustic characteristics of the material. The material attenuation curve is fitted by the least squares method to obtain the mathematical expressions of the gain compensation coefficient and frequency compensation curve;

[0104] Adopt intelligent optimization methods, such as genetic algorithms, with noise reduction effect and voice quality as objective functions, to find the optimal noise reduction parameter combination and achieve adaptive optimization of noise reduction performance;

[0105] In the process of noise reduction algorithm processing, the environmental noise and temperature and humidity sensors inside the power equipment box are used to collect noise and environmental parameter data in real time. A convolutional neural network is used to build a noise-material-environment multi-parameter mapping model. The noise spectrum, material properties, temperature and humidity and other parameters are used as input features to train the model to predict the change trend of material acoustic characteristics under different working conditions.

[0106] The cross entropy loss function and Adam optimizer are used to train the model and tune hyperparameters. The learning curve and confusion matrix are used to evaluate the generalization ability and classification accuracy of the model, and the leave-one-out method is used for model testing and performance verification. The trained model is deployed to the embedded system to predict the acoustic characteristic parameters of the material in real time and guide the dynamic parameter adjustment strategy of the noise reduction algorithm.

[0107] In view of the complex structure and diverse materials of the power equipment box, the partition modeling method is adopted to divide the interior of the box into several acoustic sub-areas according to factors such as geometric dimensions, material properties, and functional requirements; acoustic continuity conditions are imposed on the coupling surfaces between sub-areas to ensure the continuity of sound pressure and vibration velocity; applicable acoustic propagation models and noise reduction algorithms are applied in each sub-area, such as using statistical energy analysis in the metal shell area and equivalent material parameters in the sound absorption material area;

[0108] The partition modeling method can reduce the computational complexity of large acoustic systems and improve the efficiency and accuracy of numerical simulations. At the same time, by evaluating the noise reduction effect of each sub-area, the noise reduction performance of the entire power equipment box can be optimized;

[0109] Through the above technical solutions, we can accurately analyze the attenuation and distortion characteristics of different materials on audio signals based on the complex material and structural characteristics inside the power equipment box, and establish a multi-material acoustic propagation model;

[0110] According to the results of acoustic characteristics analysis, the noise reduction algorithm parameters are dynamically optimized to achieve self-adaptation to environmental changes. A noise-material-environment mapping model is established using machine learning methods to intelligently predict the trend of changes in material acoustic characteristics and guide the dynamic adjustment of noise reduction strategies.

[0111] Here is an example:

[0112] A three-dimensional acoustic model of the power equipment box was established in ANSYS finite element software. The tetrahedral mesh was used, the mesh size was set to 1 / 10 of the minimum acoustic wavelength, the boundary condition was set to impedance boundary, and the input air density was 1.225 kg / m 3 , the speed of sound is 340m / s. Material property parameters include stainless steel density 7930kg / m 3 , Poisson's ratio 0.3, sound-absorbing cotton density 32kg / m 3 , flow resistance coefficient 8000Pa·s / m 2 ;

[0113] Through harmonic response analysis, the sound pressure level response curves of different material areas in the frequency range of 20Hz to 20kHz were obtained, and the natural frequency of the stainless steel plate was extracted to be around 125Hz and 1000Hz, while the sound-absorbing cotton had obvious attenuation on the noise in the frequency band above 1000Hz.

[0114] The above data were entered into the material acoustic property database, and through multiple regression analysis, the empirical formula between material thickness, density and acoustic attenuation coefficient was obtained, such as Y = 0.021X1 + 0.078X2-1.25, where Y is the attenuation coefficient, X1 is the material thickness, and X2 is the density. In the ray acoustic model, the sound source power is set to 1W, the number of rays is 10,000, the maximum number of reflections is 5, and the receiving radius is 0.01m;

[0115] Through ray tracing calculation, the sound pressure level distribution cloud map inside the box is obtained. It can be seen that the sound pressure level at the top plate is as high as 85dB, while the sound pressure level at the sound-absorbing cotton drops to below 60dB.

[0116] The calculation results of the ray acoustics method were compared with the measured data. The average relative error was less than 3%, which met the requirements of engineering applications. In the design of the adaptive FIR filter, according to the material property database, the cutoff frequency of the stainless steel area was set to 1000Hz, the filter order was 128, and the cutoff frequency of the sound-absorbing cotton area was set to 500Hz, and the order was 64.

[0117] By analyzing the distortion and attenuation characteristics of the audio signal, the genetic algorithm is used to optimize the filter coefficients. The population size is set to 50, the evolutionary number is 100, the crossover probability is 0.8, and the mutation probability is 0.1. The optimal filter gain compensation curve is G(f)=1.2e^(-0.001f), and the frequency compensation curve is T(f)=0.023f2-0.18f+1.05. The optimized FIR filter is embedded in the DSP hardware circuit to perform adaptive noise reduction on the collected noise signal in real time;

[0118] The convolutional neural network model is trained using noise spectrum, material thickness, temperature and humidity parameters. The network structure includes 3 convolutional layers, 2 pooling layers and 2 fully connected layers. The convolution kernel size is 3x3, the step size is 1, and the activation function is ReLU.

[0119] Using the Adam optimizer, the initial learning rate is set to 0.01, the batch size is set to 32, and after training for 100 epochs, the material recognition accuracy on the test set reaches more than 95%;

[0120] The K-means clustering algorithm is used to acoustically partition the power equipment box. According to the clustering results, the box is divided into five sub-areas. The corresponding acoustic propagation model and noise reduction algorithm are applied in each sub-area, and the acoustic continuity condition is imposed at the area boundary. The sound field distribution of each sub-area is jointly solved to obtain the noise reduction effect evaluation result of the entire power equipment box, and the noise reduction amount reaches more than 20dB(A).

[0121] Furthermore, in this embodiment, step S3 specifically includes:

[0122] Environmental noise sensors are set at different locations in the power equipment box to collect noise data samples under different working conditions and form a noise sample database; the environmental noise sensor can be a MEMS microphone array or a capacitive sound meter;

[0123] The signal processing method is used to perform time-frequency domain analysis on the collected noise sample data, extract the spectrum characteristics, energy characteristics and statistical characteristics of the noise signal, and construct a noise feature vector data set; the noise samples include background noise, equipment operation noise and abnormal noise, and the signal processing method can select wavelet transform and Fourier transform. The spectrum characteristics, energy characteristics and statistical characteristics of the noise signal include frequency band energy ratio, spectral entropy, spectral centroid and kurtosis;

[0124] Using machine learning algorithms, combined with noise feature vector data sets, training machine learning models, classifying and identifying different types of noise, obtaining noise type labels and intensity feature information, and forming a noise identification knowledge base; machine learning algorithms can be selected from support vector machines (SVM), K nearest neighbors (KNN), and decision trees;

[0125] Aiming at the complex acoustic environment inside the power equipment box, a convolutional neural network (CNN) and a long short-term memory network (LSTM) based on the ResNet or Inception network structure are used to establish an end-to-end deep learning model for environmental noise classification and recognition.

[0126] The original waveform data of the noise signal is normalized and preprocessed with data enhancement. Through feature learning and classification prediction, the model is trained using the cross entropy loss function and the Adam optimization algorithm. The model hyperparameters are tuned using grid search and Bayesian optimization methods to improve the accuracy and generalization ability of noise recognition.

[0127] The prior information such as noise type and intensity in the noise identification knowledge base is integrated with environmental parameters (such as temperature and humidity) and equipment operating parameters (such as speed and current), and a multi-dimensional noise scene feature space is constructed based on feature selection principles and quantification methods.

[0128] Adopting the Bayesian network probability graph model, designing a reasonable network structure, learning the conditional probability table through maximum likelihood estimation or Bayesian estimation method, establishing the mapping relationship between noise scenes and noise reduction strategies, and realizing adaptive recognition and noise reduction processing of noise environments;

[0129] In the process of adaptive noise reduction, reinforcement learning algorithms such as DQN or DDPG are used to define appropriate state space, action space and reward function. Through Q-value function approximation and exploration and utilization of balance strategy, combined with experience replay and target network update mechanism, the decision threshold of the noise recognition model and the parameter setting of the noise reduction algorithm are dynamically adjusted according to the feedback information of the noise reduction effect, so as to continuously optimize the performance of the noise reduction system and improve the real-time and accuracy of noise reduction.

[0130] By continuously collecting environmental noise data inside the power equipment box, using online learning and incremental learning algorithms such as online gradient descent and incremental support vector machine, designing a balance strategy between old knowledge retention and new knowledge learning, and regularly updating and expanding the noise identification knowledge base and prior information;

[0131] Establish an evaluation index system for the performance of noise identification models and set a trigger mechanism for model updates so that the noise reduction system can adapt to changes in the noise environment and maintain a long-term stable noise reduction effect, providing reliable acoustic information support for operation monitoring and fault diagnosis of power equipment;

[0132] Here is an example:

[0133] A 64-channel MEMS microphone array is installed inside the power equipment box, with a sampling frequency of 48kHz and a bit depth of 24bit;

[0134] The collected noise data was decomposed by wavelet packet, and Daubechies5 wavelet basis function was used to decompose 5 layers, and 12 features such as energy value, standard deviation, and kurtosis of each sub-band were extracted. The spectrum was analyzed by 1 / 3 octave band, and 6 features such as energy ratio and band centroid of 27 octave bands were calculated;

[0135] Three features, including short-time energy and zero-crossing rate, were extracted from the time domain waveform. A total of 21-dimensional noise feature vectors were constructed. The K-means clustering algorithm was used to perform cluster analysis on the feature data, and the optimal number of clusters was determined to be 5 by the silhouette coefficient.

[0136] The clustering results were classified and learned through the SVM model, and the radial basis kernel function was used. The penalty factor C and the kernel function parameter g were optimized through grid search, and the 5-fold cross validation accuracy reached 92%. A 3-layer CNN model was built, with a convolution kernel size of 3x3, a pooling layer of 2x2MaxPooling, and an activation function of ReLU;

[0137] Input the normalized noise waveform image, Adam optimizer, initial learning rate 0.001, min i-batch size 64, training 100 epochs, the classification accuracy on the test set reached 96%;

[0138] Integrate 8-dimensional environmental and operating parameters such as temperature, humidity, current, and speed to build a Bayesian network, learn the network structure through the maximum weight spanning tree algorithm, estimate parameters through the EM algorithm, judge the noise scene, and use MAP to infer the noise reduction strategy. Define the state space of the noise reduction task as 18-dimensional noise features, the action space as the noise reduction filter parameters, and the reward function as the improvement in signal-to-noise ratio;

[0139] Adopt DDPG algorithm, Actor network 2-layer 128-node LSTM, Cr it ic network 3-layer 256-node MLP. Set the experience noise to 0.1, discount factor to 0.95, experience replay cache to 1,000,000 records, and target network soft update factor to 0.01; train 2,000 epi sodes, and the average return increases from 2dB to 15dB;

[0140] The online noise data uses moving average to update the noise type recognition model with an update weight of 0.1, and retraining is triggered when the classification accuracy is lower than 90%.

[0141] Furthermore, in this embodiment, step S4 specifically includes:

[0142] Combining the characteristics of environmental noise, acoustic properties of materials and properties of different locations, the corresponding noise reduction algorithms and parameters are used to perform adaptive noise reduction processing on audio signals.

[0143] Adaptive noise reduction processing of audio signals includes eliminating noise interference and retaining humidity volume feature information;

[0144] Specifically, the environmental noise signal inside the power equipment box is obtained, and the frequency domain, time domain and time-frequency domain characteristics of the noise signal are extracted through signal processing technology, such as spectrum analysis and wavelet transform. The frequency domain, time domain and time-frequency domain characteristics of the noise signal include power spectrum density, MEL cepstral coefficients and wavelet packet energy to form an environmental noise feature vector;

[0145] According to the structural layout and material property information of the power equipment box, the finite element method is used to calculate and analyze the acoustic characteristic parameters of different materials, obtain the sound absorption coefficient, reflection coefficient and acoustic impedance parameters of the material, and establish a material acoustic characteristic database;

[0146] Combining factors such as environmental noise characteristics, material acoustic properties, and sensor location, an acoustic propagation model is constructed. Numerical simulation methods such as ray tracing and boundary element method are used to calculate the propagation attenuation characteristics and frequency response curves of noise at different locations inside the power equipment box, forming a position attribute mapping relationship.

[0147] Taking the environmental noise characteristics, material acoustic properties and location attributes as input feature vectors, supervised machine learning algorithms such as support vector machines, random forests and XGBoost are used to train with a large number of noise samples and denoising algorithm selection results to establish a denoising algorithm selection model.

[0148] The output of the model is the type and parameters of the optimal denoising algorithm. The model hyperparameters are optimized through methods such as cross-validation and grid search to improve the accuracy and generalization ability of algorithm selection. Intelligent optimization algorithms, such as Bayesian optimization and particle swarm optimization, are used to automatically search for the optimal parameter combination of the selected denoising algorithm.

[0149] Combining multiple performance indicators such as signal quality after noise reduction, noise suppression level, and humidity and volume feature retention, we designed a suitable objective function and introduced a subjective evaluation mechanism through manual listening and scoring to guide the convergence direction of the optimization algorithm.

[0150] According to the changes in environmental noise and the feedback on the noise reduction effect, the optimization results are updated online, and the algorithm parameters are adjusted in real time to achieve adaptive optimization;

[0151] Apply the automatically selected noise reduction algorithm and optimized parameters to the audio signal processing process to perform adaptive noise reduction on the collected noise signal;

[0152] A regularization term for preserving humidity and volume features is added to the cost function of the adaptive filter to guide the filter to minimize the distortion of humidity and volume features while reducing noise.

[0153] The NSGA-II multi-objective optimization algorithm is used to take noise reduction performance and humidity volume feature retention as two optimization objectives, and to find the Pareto optimal solution set of the two.

[0154] The evaluation index of humidity volume feature retention is designed, including the signal-to-noise ratio and correlation coefficient of humidity volume feature, as one of the performance evaluation criteria of adaptive noise reduction, to achieve maximum elimination of noise interference and simultaneous retention of humidity volume feature.

[0155] Here is an example:

[0156] 20 MEMS microphones are arranged on the 8 corners and 12 edges of the power equipment box, with a sampling frequency of 48kHz and a bit depth of 24bit;

[0157] Perform 1024-point FFT transformation on the collected environmental noise signal to extract the power spectrum density characteristics in the 0-20kHz frequency band. Use the 10th-order Daubechies wavelet to perform 5-layer decomposition on the signal to extract 12 characteristic parameters such as energy, standard deviation and kurtosis of each sub-band;

[0158] A 39-order Mel filter bank is used to extract MFCC features, including 12 static coefficients and first-order and second-order differential coefficients. The acoustic transmission matrix method is used to calculate the sound absorption coefficient, reflection coefficient and characteristic impedance of five materials inside the box, including steel plate, aluminum plate, and foam sound-absorbing material. The frequency range is 50Hz-10kHz and the frequency interval is 1 / 3 octave.

[0159] The ray acoustic method is used to simulate the propagation of environmental noise interference inside the box. 1000 sound rays are emitted, the maximum number of reflections is 10 times, and the frequency range is 50Hz-10kHz. The sound pressure level frequency response curve and transfer function at 20 sensor positions are calculated, and a noise-material-position attribute data set with 180 features is constructed. The LI BSVM toolbox is used to train the SVM classifier, RBF kernel function, 5-fold cross validation, and grid search to obtain the optimal parameters C=10, g=0.1, and the classification accuracy reaches 95%;

[0160] The particle swarm optimization algorithm is used to optimize the order, step size and leakage factor of the LMS adaptive filter. The objective function is the weighted sum of the noise reduction signal-to-noise ratio and the correlation coefficient of the humidity and volume characteristics.

[0161] 20 particles were initialized, the maximum number of iterations was 50, the inertia weight was 0.7, the individual learning factor and the social learning factor were both 1.5, and the optimal parameter combination was selected through subjective evaluation experiments and used for adaptive denoising; multi-objective optimization of denoising and humidity and volume feature retention was achieved based on the NSGA-II algorithm, with a crossover probability of 0.8, a mutation probability of 0.1, a population size of 50, and 100 generations of evolution; the average denoising SNR of the Pareto optimal solution set was improved by 10dB, and the average humidity and volume feature correlation coefficient reached above 0.95.

[0162] Furthermore, in this embodiment, step S5 specifically includes:

[0163] For the audio signal after adaptive noise reduction, the time-frequency analysis method is used to extract the volume characteristic parameters reflecting the humidity state to form a high-dimensional humidity volume characteristic vector; the volume characteristics include short-time energy, logarithmic energy, zero-crossing rate, frequency band energy ratio, spectral centroid, spectral entropy and other time-frequency domain characteristics; the time-frequency analysis method can select short-time Fourier transform;

[0164] The audio signal is analyzed at multiple scales and multiple resolutions through wavelet packet decomposition, and wavelet coefficient statistics are extracted at different frequency bands and time scales to capture the non-stationary humidity volume characteristics in the audio signal; the wavelet coefficient statistics include humidity-sensitive energy, standard deviation, and kurtosis;

[0165] The speech feature extraction method is used to obtain the spectrum distribution characteristics reflecting the humidity state, and the dynamic characteristics of volume intensity changing with humidity are obtained; the speech feature extraction method selects Mel frequency cepstrum coefficient (MFCC); the spectrum distribution characteristics include MFCC coefficients, first-order difference coefficients and second-order difference coefficients;

[0166] The correlation and significance between each humidity volume feature and humidity state are evaluated through feature selection algorithm;

[0167] The maximum correlation minimum redundancy algorithm is used to first select the feature with the greatest correlation with the humidity state, and then select the feature with the least correlation with the selected feature from the remaining features until the preset number of features or correlation threshold is met;

[0168] The chi-square test was used to evaluate whether the feature and humidity status were independent of each other by calculating the difference between the actual observed frequency and the expected frequency of the feature value;

[0169] For the continuous humidity and volume features, it is necessary to discretize them first, then calculate the chi-square value, and finally select several features with the largest mutual information value or the largest chi-square value as the optimal feature subset to reduce the feature dimension and calculation complexity;

[0170] The feature dimension reduction techniques such as principal component analysis (PCA) and independent component analysis (ICA) are used to further compress the dimension of the feature vector while retaining the essential information of humidity and volume features.

[0171] For the PCA algorithm, the humidity volume feature data is first centered and standardized, and then the feature covariance matrix is ​​calculated, and its eigenvalues ​​and eigenvectors are solved. The eigenvectors corresponding to the first k largest eigenvalues ​​are selected to form a transformation matrix, and the original high-dimensional features are mapped to a low-dimensional space. The k value can be determined based on the cumulative contribution rate or reconstruction error;

[0172] For the ICA algorithm, it is assumed that the observed humidity volume feature is a linear mixture of several independent source signals. By optimizing the objective function of maximum likelihood estimation or minimum mutual information, an unmixing matrix is ​​estimated to restore the mixed features to independent source signal components as the low-dimensional feature vector after dimensionality reduction;

[0173] After normalizing the multi-domain humidity volume features, feature fusion is performed using strategies such as feature-level fusion or decision-level fusion, which comprehensively utilizes the characterization capabilities of different features for humidity states. In feature-level fusion, feature vectors from different domains can be directly concatenated into a longer feature vector, or methods such as multi-kernel learning can be used to learn the optimal weighted combination of different feature subsets.

[0174] In decision-level fusion, a humidity estimation sub-model can be trained for each feature domain separately, and then the estimation results of each sub-model can be integrated through weighted averaging, voting and other strategies to obtain the final humidity estimation value. Construct a humidity estimation model, take the fused humidity volume features as input, train and optimize the model parameters, and establish a nonlinear mapping relationship between features and humidity values;

[0175] For the support vector regression (SVR) model, the Gaussian kernel is selected as the kernel function, and the grid search method is used to optimize the penalty factor C, the number of support vectors and the kernel function width parameter. The mean square error (MSE) is used as the evaluation index to select the parameter combination with the best generalization performance.

[0176] For the Gaussian process regression (GPR) model, the square exponential kernel is selected as the covariance function, and the maximum likelihood estimation method is used to learn the prior mean function and noise variance and other parameters. The negative logarithmic marginal likelihood is used as the loss function, and the gradient descent method is used for optimization. In the prediction stage, the GPR model can give the mean of humidity estimation and estimate the uncertainty of humidity. Finally, the performance of the humidity estimation model is evaluated on the test set, and the mean absolute error, root mean square error and other indicators are calculated to optimize the model structure and parameters, thereby improving the accuracy and robustness of humidity estimation.

[0177] Here is an example:

[0178] Perform short-time Fourier transform on the denoised audio signal, use the Hanning window function, with a window length of 20ms and a frame shift of 10ms, calculate the 256-point FFT of each frame, and extract the 96-dimensional MFCC features of the 32-dimensional MEL filter group energy spectrum, first-order difference spectrum, and second-order difference spectrum;

[0179] The signal was decomposed by 5-layer wavelet packet, and the Daubechies-4 wavelet basis was used to extract 15-dimensional wavelet domain features such as energy, standard deviation and standardized center distance of each subband. The mutual information value of volume feature and humidity was calculated using the mRMR algorithm, and the maximum correlation number was set to 10 and the minimum redundancy number was set to 5. Redundant features were iteratively deleted until the 30-dimensional optimal feature subset was retained;

[0180] The PCA algorithm is used to decompose the feature matrix, and the first five principal components are selected, with a cumulative contribution rate of more than 95%, to achieve feature dimensionality reduction to 5 dimensions. The MFCC, wavelet domain, and time domain features are Z-score normalized, and then linearly fused with a weight ratio of 1:2:1 to obtain a 12-dimensional humidity estimation feature vector;

[0181] Train the SVR humidity estimation model, taking the Gaussian kernel function as an example, and determine the penalty factor C as 10 through 5-fold cross validation, the kernel function width parameter g as 0.1, and the number of support vectors as 30% of the number of training samples;

[0182] On the test set, the average absolute error between the humidity estimate and the actual value is 3%, the root mean square error is 0.05, and the Pearson correlation coefficient is above 0.95.

[0183] Furthermore, in this embodiment, step S6 specifically includes:

[0184] The audio signals of different materials in different positions of the power equipment box are obtained through adaptive noise reduction, the volume characteristic parameters reflecting the humidity state are extracted, and a multi-dimensional humidity volume characteristic vector is constructed to form the input data set of the humidity estimation model; the volume characteristic parameters include root mean square energy, frequency band energy ratio, and Mel cepstral coefficient;

[0185] Among them, the root mean square energy is extracted by calculating the root mean square value of the samples of each frame of the audio signal, reflecting the overall energy level of the volume; the band energy ratio divides the audio signal into several bands, calculates the proportion of the energy in each band to the total energy, and describes the frequency domain distribution characteristics of the volume; the Mel cepstral coefficient is obtained by mapping the spectrum of the audio signal to the Mel scale and then performing a discrete cosine transform, reflecting the perceptual characteristics of the volume;

[0186] A long short-term memory neural network is used to build a deep learning humidity estimation model, and a Dropout layer is added between the LSTM layer and the fully connected layer to prevent the model from overfitting;

[0187] The mean square error is selected as the loss function of the humidity estimation model, and the Adam optimizer is used to adaptively adjust the model parameters. The model parameters with the lowest mean square error during the training process are saved as the optimal model.

[0188] The humidity volume feature data of different materials at different locations are divided into training set, validation set and test set by stratified sampling method. According to the distribution of humidity labels, the data set is divided into multiple humidity intervals, and then samples are randomly selected in each interval at a ratio of 8:1:1 to ensure the balance and representativeness of the data set division;

[0189] Adopt the method of multiple random partitioning, repeat the data set partitioning K times, train and evaluate the model each time, and finally take the average of the K results as the final performance indicator of the model to reduce the random impact of data set partitioning;

[0190] The training set is used to train the deep learning humidity estimation model, the validation set is used to evaluate the model performance and tune the parameters, and the test set is used to test the generalization ability of the model;

[0191] During the training process, the hyperparameters of the LSTM layer are dynamically adjusted to optimize the model's fitting ability and generalization performance. At the same time, an early stopping mechanism is introduced to terminate the training in advance to prevent the model from overfitting. The hyperparameters of the LSTM layer include the number of neurons and the dropout probability.

[0192] Deploy the trained deep learning humidity estimation model to embedded devices or cloud servers, receive audio signals from different materials at different locations in the power equipment box in real time, extract humidity volume feature parameters, input them into the model for humidity estimation, and obtain real-time humidity values;

[0193] Furthermore, an interactive visualization interface is developed using Web front-end technologies, such as HTML, CSS, and JavaScript, to display the humidity distribution at different locations in the power equipment box in the form of a heat map, with different colors representing different humidity levels;

[0194] Use native development tools of Android or iOS platforms, such as AndroidStud io and Xcode, to develop a mobile APP to display the humidity status of the device in real time and show the humidity change trend in the form of a line graph. Set the humidity threshold control in the visual interface to allow users to customize the upper and lower limits of humidity. When the estimated humidity value exceeds the set range, an alarm prompt will be issued in the form of a pop-up window, vibration, voice, etc., and the alarm time, location and other information will be recorded to facilitate subsequent data analysis and equipment maintenance; through visualization and alarm mechanisms, real-time monitoring and early warning of humidity in the power equipment box can be achieved, and the intelligent level of equipment operation and maintenance can be improved;

[0195] Here is an example:

[0196] The audio signals collected from 10 different locations in the power equipment box were preprocessed, and the wavelet threshold denoising method was used to set the threshold of the wavelet coefficient to 3 times the standard deviation to remove high-frequency noise.

[0197] 12-dimensional MFCC features, 39-dimensional spectral features, 1-dimensional short-time energy and 1-dimensional zero-crossing rate were extracted to form a 53-dimensional humidity volume feature vector. Based on the humidity label, the data set was divided into three intervals: 0-30%, 30%-60%, and 60%-100%. In each interval, 80% was randomly selected as the training set, 10% as the validation set, and 10% as the test set. The division was repeated 10 times, and the average of the 10 evaluation results was taken.

[0198] Construct a 3-layer LSTM network and a 2-layer fully connected network. The LSTM layer contains 256, 128, and 64 units, and the Dropout probability is 0.3. The fully connected layer contains 128 and 1 units. Use the Adam optimizer, with an initial learning rate of 0.005, a dynamic learning rate decay factor of 0.8, a training batch size of 64, and a maximum number of iterations of 200. Introduce an early stopping mechanism to stop training when the validation set loss does not decrease for 20 consecutive iterations to prevent overfitting.

[0199] The model performance was evaluated on the test set, with an average absolute error of 2.5% and a root mean square error of 3.2%. A visual web application was developed using the Vue.js framework, which receives the estimated humidity values ​​of 10 measurement points in real time through WebSocket, refreshes once a second, and displays the humidity distribution inside the box in a 10x10 heat map matrix, with a color code mapping the humidity range. An iOS APP was developed using the Swift language, which obtains humidity data through the HTTP protocol and displays the humidity change curve for the past hour in real time as a line chart;

[0200] The humidity warning threshold is set to 70%. When the threshold is exceeded, a warning box pops up and a voice reminder is given. At the same time, warning data is sent to the cloud, including warning time, location, and humidity value, to facilitate timely processing by operation and maintenance personnel.

[0201] Furthermore, in this embodiment, step S7 specifically includes:

[0202] If the humidity estimation value of a certain location in the power equipment box output by the deep learning humidity estimation model is higher than the preset humidity threshold, a humidity abnormality alarm message is automatically generated, the time, location and humidity value of the abnormality are recorded, and the automatic dehumidification device at that location is triggered to start, and the corresponding location in the power equipment box is dehumidified until the humidity returns to the theoretical normal humidity value;

[0203] Further, according to the ambient temperature monitoring data of the location where the humidity anomaly occurs, a quantitative relationship model between the ambient temperature and the dehumidification efficiency is established by analyzing the historical data, such as a linear regression model or a nonlinear model;

[0204] The established model is used to predict the optimal combination of dehumidification parameters, such as refrigerant flow and compressor power. During the dehumidification process, the ambient temperature and humidity data are collected in real time. The model parameters are dynamically adjusted through the optimization algorithm to achieve online adaptive optimization of the dehumidification parameters, thereby minimizing the dehumidification energy consumption while ensuring the dehumidification effect.

[0205] After the automatic dehumidification device is started, the humidity change data of the location is obtained in real time, and the dehumidification process is continuously monitored through humidity sensors and deep learning humidity estimation models;

[0206] According to the humidity change data collected by the humidity sensor, the humidity drop rate and change trend, such as the first-order derivative and the second-order derivative, are calculated;

[0207] Establish a mapping relationship between humidity change characteristics and the working state of the dehumidification device, such as fuzzy control rules or decision tree models, and dynamically adjust the working state of the dehumidification device, including compressor frequency and fan speed, according to the humidity change characteristics calculated in real time and using the established mapping relationship to achieve optimal control of dehumidification efficiency;

[0208] If the humidity decreases significantly slower than expected during the dehumidification process, or the humidity cannot be reduced below the threshold for a long time, it is judged that the automatic dehumidification device may be faulty;

[0209] Based on historical operation data, summarize the common failure modes and characteristics of dehumidification devices, such as refrigerant leakage, compressor overload, etc., and establish a failure knowledge base;

[0210] During the dehumidification process, the operating parameters of the dehumidification device, including current, voltage, and temperature, are collected in real time to extract fault-related feature quantities;

[0211] Adopting knowledge-based reasoning methods, such as fault tree analysis, to match the extracted fault features with the fault patterns in the knowledge base, to achieve rapid fault diagnosis and location, generate fault alarm information, indicate possible fault causes and treatment measures, and guide operation and maintenance personnel to carry out targeted maintenance;

[0212] When the humidity monitoring data and estimated values ​​are lower than the preset threshold for multiple consecutive times, it is judged that the humidity has returned to normal levels, and the dehumidification device is automatically turned off to save energy. At the same time, the dehumidification record of the location is updated, including the start and end time of dehumidification and the humidity change curve, which is convenient for subsequent analysis and optimization;

[0213] After the automatic dehumidification device is turned off, the humidity status of the location continues to be monitored. If the humidity exceeds the threshold again within a short period of time, the automatic dehumidification device is restarted and the dehumidification time is appropriately extended to prevent repeated fluctuations in humidity.

[0214] Here is an example:

[0215] By performing multiple linear regression analysis on 100 sets of historical temperature and dehumidification efficiency data, the linear relationship between ambient temperature and compressor power and refrigerant flow rate was obtained, with the residual square sum less than 0.1 and the goodness of fit reaching more than 95%;

[0216] The estimated humidity value is compared with the preset humidity threshold. If the humidity exceeds the threshold, the automatic dehumidification device is triggered to start and dehumidify the power equipment box until the humidity returns to the theoretical normal humidity value;

[0217] When the deviation between the real-time collected temperature and humidity data and the fitting model exceeds 5%, the parameter adaptive adjustment is triggered, and the regression coefficient is updated by the gradient descent method with a learning rate of 0.01. The mean square error is reduced to below 0.05 after 20 iterations.

[0218] During the dehumidification process, the first-order and second-order differences of humidity are calculated every 30 seconds, representing the rate of decline and acceleration, respectively. The humidity change rate is mapped to the dehumidification intensity level (level 1-5), and the fuzzy control algorithm is used to output the fuzzy control quantity of the compressor frequency (20-100Hz) and the fan speed (500-3000rpm). The maximum membership method is used to defuzzify and realize the dynamic adjustment of the dehumidification device.

[0219] Extract characteristic vectors of 10 typical faults from the fault knowledge base, including 15 parameters such as compressor current, refrigeration temperature, and fan vibration;

[0220] The fault diagnosis method based on case reasoning is adopted to calculate the Euclidean distance between the currently collected operating parameters and each fault feature vector. The fault type with the smallest distance is taken as the diagnosis result. If the minimum distance is greater than a threshold (such as 1.5), it is judged as an unknown fault.

[0221] After the dehumidification is completed, the humidity change curve of this dehumidification process is extracted, and the dynamic time warping (DTW) algorithm is used to match the similarity with the historical dehumidification process. The process parameters with the highest similarity are used as the initial parameters for the next dehumidification, and fine-tuned according to indicators such as the dehumidification duration and energy consumption to achieve dehumidification optimization control.

[0222] Furthermore, in this embodiment, step S8 also includes:

[0223] According to the humidity changes at different locations in the power equipment box, the working parameters of the automatic dehumidification device are adjusted in real time, including the dehumidification intensity and duration;

[0224] Specifically, multiple high-precision humidity sensors are arranged at different locations inside the power equipment box, including the top, bottom, and sides, to collect humidity data at each location in real time;

[0225] The humidity sensor sends humidity data to the controller node through the Zigbee wireless communication protocol, and then the controller node transmits it to the dehumidification control system through the RS485 bus or Ethernet. After receiving the humidity data of each location, the dehumidification control system performs denoising, outlier filtering and other operations through the data preprocessing module to obtain stable and reliable humidity change information, which is stored in the humidity database and triggers the dehumidification strategy analysis module at the same time;

[0226] The dehumidification strategy analysis module reads humidity change data within a certain time range from the humidity database, and predicts the humidity change trend of each location in the future period of time through time series analysis algorithms such as ARIMA model or Kalman filter;

[0227] Among them, the ARIMA model establishes a dynamic regression equation for humidity change by identifying the parameters p, d, and q of the three parts of autoregression (AR), difference (I), and moving average (MA); the Kalman filter dynamically tracks the real change of humidity by establishing the state transfer equation and observation equation, recursively calculating the prior estimate and the posterior estimate of the state variable;

[0228] When applying these algorithms, the humidity time series is first tested for stationarity and white noise, and differential operations or detrending are performed when necessary. The maximum likelihood estimation or least squares method is used to fit the parameters, and the model is optimized according to criteria such as AIC and BIC, and finally the humidity change rate, humidity gradient and other parameters are obtained.

[0229] Based on the predicted humidity change parameters and the preset target humidity range, the dehumidification strategy analysis module calculates the dehumidification intensity and duration required for each location, generates a dehumidification strategy matrix, and sends it to the dehumidification device control module;

[0230] The dehumidification device control module analyzes the dehumidification strategy matrix to obtain the working parameter settings of the dehumidification devices at different locations, such as compressor power, fan speed, and dehumidifier dosage. These parameters are sent to each dehumidification device through control signals to achieve differentiated dehumidification control.

[0231] During the dehumidification process, humidity sensors continuously monitor humidity changes at various locations. At the same time, temperature sensors are arranged near the dehumidification device to collect temperature data at different locations and upload them to the dehumidification control system together with humidity data.

[0232] The dehumidification control system uses spatial interpolation algorithms such as Kriging interpolation and inverse distance weighted interpolation to construct continuous temperature and humidity field distribution surfaces based on discrete temperature and humidity measurement point data. By performing gradient analysis and contour line extraction operations, the high and low temperature and humidity areas, change trends, etc. are determined, providing a decision-making basis for collaborative optimization of multi-point dehumidification.

[0233] For example, in areas with higher temperatures and lower humidity, the dehumidification intensity can be appropriately reduced; in areas with lower temperatures and higher humidity, the dehumidification intensity can be appropriately increased;

[0234] At the same time, computational fluid dynamics (CFD) methods, such as finite element analysis and finite volume method, are used to simulate the dynamic changes of temperature and humidity fields during the dehumidification process, optimize the layout and working parameters of the dehumidification device, and improve the overall dehumidification efficiency. When the real-time humidity data is fed back to the dehumidification control system, the dehumidification strategy analysis module dynamically tracks the humidity changes. If it is found that the actual humidity changes deviate significantly from the predicted trend, the dehumidification strategy is adjusted in time, a new dehumidification strategy matrix is ​​generated, and the dehumidification device control module is notified to adjust the working parameters of the dehumidification device accordingly;

[0235] Here is an example:

[0236] A total of 14 SHT31 temperature and humidity sensors are arranged in the eight corners and the center of the six faces of the power equipment box. The CC2530 chip forms a Zigbee network and uploads data every 10 seconds at a rate of 250kbps.

[0237] The dehumidification control system filters the humidity data of each sensor through the Ka lman filtering algorithm. The mean square error of the humidity estimation value is less than 1%RH. Then the ARIMA (2,1,1) model is used to predict the humidity trend. The residual sum of squares of the model fitting is less than 0.05, and the average relative error of the prediction is less than 5%.

[0238] According to the predicted humidity change rate and humidity distribution gradient, the dehumidification strategy matrix of 14 locations in the next hour is calculated. When the humidity rise rate exceeds 3%RH / min or the humidity difference between different locations exceeds 10%RH, the dehumidification intensity is increased by 20% and the dehumidification time is extended by 15 minutes.

[0239] The dehumidification strategy matrix is ​​converted into specific hardware control parameters, such as increasing the compressor power by 20%, adjusting the corresponding frequency from 50Hz to 60Hz, increasing the fan speed from 3000rpm to 3600rpm, and increasing the dehumidifier dosage from 100ml / h to 120ml / h;

[0240] By performing Kriging interpolation on the data of 14 temperature and humidity measurement points, a 100x100 temperature and humidity field distribution matrix was generated, and three high-temperature areas (temperature>35℃) and two high-humidity areas (humidity>85%RH) were identified. The dehumidification intensity was increased by 30% in these areas, and the dehumidification duration was extended by 30 minutes. The CFD software ANSYS Fluent was used to numerically simulate the dehumidification process, using the k-eps i lon turbulence model and VOF multiphase flow model, with a grid number of about 500,000 and a calculation step size of 0.1s. The simulation results showed that the optimized dehumidification efficiency increased by 15% and the energy consumption decreased by 10%;

[0241] During actual operation, if the deviation between the real-time humidity data and the predicted value exceeds 10%, the dehumidification strategy is recalculated and the updated control parameters are issued, so that 14 dehumidification devices work together to concentrate dehumidification in 3 high-temperature areas and 2 high-humidity areas. After 15 minutes, the overall humidity in the box drops to below 70% RH and the maximum temperature drops to below 30°C, meeting the environmental requirements for safe operation.

[0242] Furthermore, in this embodiment, step S9 also includes:

[0243] If no abnormality is detected, the original dehumidification working state is maintained;

[0244] If an abnormal situation is detected, the dehumidification work will be suspended, the alarm mechanism will be triggered, and an alarm message and data recording the abnormal situation will be sent;

[0245] Specifically, in the dehumidification control system, a judgment threshold for humidity exceeding the standard and dehumidification device failure is set. The dehumidification control system monitors the humidity data of each location and the operation status data of the dehumidification device in real time, and judges whether there are abnormal conditions such as humidity exceeding the standard or dehumidification device failure by comparing with the preset threshold value;

[0246] When no abnormality is detected, the original dehumidification working state is maintained;

[0247] When an abnormal situation is detected, the dehumidification control system immediately triggers the alarm mechanism, pops up a warning prompt on the display screen, and sends an alarm message to the on-duty personnel through SMS, email, APP push, etc., informing them of the time, location, type and severity of the abnormality.

[0248] The dehumidification control system automatically generates an abnormal event report, which records in detail the start and end time of the abnormality, duration, humidity change curve, operating parameters of the dehumidification device and other data, and saves the report in the database and sends it to relevant technical personnel at the same time;

[0249] The abnormal report analysis module uses K-means and DBSCAN clustering algorithms to automatically group abnormal data, identify different types of abnormal patterns, perform statistical analysis on each abnormal category, and calculate its occurrence frequency, average duration, humidity exceeding the standard range and other indicators;

[0250] Using April and FP-growth association rule mining algorithms, we can discover the associations and time series rules between different anomaly categories. Through the visualization of anomaly data, such as scatter plots, parallel coordinates graphs and other charts, we can intuitively present the distribution characteristics and evolution trends of anomalies.

[0251] The abnormal report analysis results are passed to the fault diagnosis module, which builds a fault diagnosis model based on the machine learning algorithm of decision tree, random forest or support vector machine, takes the abnormal data features as input and the fault type as output, and uses k-fo ld cross validation and grid search methods to train and optimize the model. At the same time, the knowledge base is introduced to perform rule-based verification and supplement the diagnosis results of the machine learning model, forming a hybrid diagnosis system combining man and machine.

[0252] In addition, integrated learning techniques, such as stacking and bagging, can be used to combine the outputs of multiple basic diagnostic models to further improve the accuracy and robustness of diagnosis;

[0253] The diagnostic results and repair suggestions generated by the fault diagnosis module are sent to the operation and maintenance management system together with the abnormal event report; the operation and maintenance management system automatically determines the priority and response time limit of the maintenance task based on the severity of the fault and the diagnostic results;

[0254] Combined with the maintenance personnel's skill level, work schedule and other factors, the Hungarian algorithm, genetic algorithm and other task allocation optimization methods are used to reasonably schedule maintenance tasks and generate maintenance work orders, which are pushed to relevant technicians in real time through APP or SMS.

[0255] Maintenance personnel inspect, maintain and repair the dehumidification device according to the work order requirements, eliminate faults and restore normal operation. During the maintenance process, the maintenance work order management system tracks the execution progress and completion status of the maintenance task, and synchronizes the feedback data during the maintenance process (such as maintenance time, replacement parts, etc.) to the anomaly detection and diagnosis system in real time to optimize the fault diagnosis model and knowledge base, continuously improve the level of intelligent operation and maintenance, and form a complete fault handling closed loop;

[0256] Here is an example:

[0257] The dehumidification control system collects humidity data and operating status parameters of the dehumidification device at 14 locations every 1 minute. If it is detected that the humidity at a certain location exceeds 90% RH for 15 minutes, or the compressor current exceeds the rated value by 30%, and the refrigerant pressure is lower than 0.2MPa for 10 minutes, it is judged as abnormal and an alarm is triggered.

[0258] Automatically send alarm information to management personnel via voice and pop-up window, notify on-duty personnel via SMS and email, and generate abnormal reports;

[0259] The abnormal report includes abnormal occurrence time, duration, humidity peak value, abnormal parameter name and deviation value;

[0260] The abnormal data were grouped using the density-based DBSCAN clustering algorithm. The eps radius parameter was set to 0.5, and the minPts minimum sample number was set to 5. The characteristics of each cluster, such as the humidity mean, duration, and abnormal parameters, were extracted to obtain three typical abnormal patterns.

[0261] Using April association rule mining algorithm, the support threshold is set to 20%, the confidence threshold is set to 60%, and it is found that the frequency of abnormal parameters A and B appearing at the same time is 75%, and the probability of C appearing within 10 minutes after B appears is 80%;

[0262] The abnormal data visualization uses the t-SNE dimensionality reduction algorithm to reduce the 54-dimensional abnormal feature vector to obtain a 2-dimensional scatter plot, and it is found that the abnormal data is mainly distributed in 3 areas. The fault diagnosis model uses the random forest algorithm, with 73 abnormal features as input and 4 fault types as output. Using 10-fold cross validation, the average accuracy of the optimal model is 93.5%. The diagnosis results are matched with the 24 fault judgment rules in the knowledge base. If they meet the rules, the diagnosis is established, otherwise further manual analysis is performed;

[0263] The maintenance task is scheduled and optimized using genetic algorithm, with real number coding, crossover probability of 0.8, mutation probability of 0.1, and a converged solution obtained after 200 iterations. The total maintenance time is shortened by 15% compared with the first-in-first-out method.

[0264] The maintenance process data is automatically entered into the work order management system through QR code scanning, and statistical analysis is performed on fault types with a diagnostic accuracy rate below 90%, optimizing the diagnostic rules and algorithms to increase the accuracy rate by 1.5 percentage points.

[0265] In the description of the present invention, it is necessary to understand that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "lateral, vertical, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the device or element referred to must have a specific direction or be constructed and operated in a specific direction. Therefore, they cannot be understood as limiting the scope of protection of the present invention.

[0266] For those skilled in the art, various other corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all of these changes and deformations should fall within the protection scope of the claims of the present invention.

Claims

1. An automatic dehumidification method in a power equipment box, characterized in that: include: S1, collecting audio signals of environmental noise at different positions in the box of the power equipment, and preprocessing the audio signals; S2. Analyze the acoustic characteristics of different materials in the power equipment box and establish an acoustic propagation model; S3. Classify and identify environmental noise at different locations and in different scenes in the power equipment box to obtain noise type and intensity characteristic information; S4, performing adaptive noise reduction processing on the audio signal based on the environmental noise characteristics, the acoustic properties of the material and the attributes of different positions; S5, extracting features from the audio signal after adaptive noise reduction to obtain key volume feature parameters reflecting humidity status; S6. Estimate the humidity level in the power equipment box according to the key volume characteristic parameters to obtain real-time humidity values ​​of different materials at different locations; S7, preset a humidity threshold, compare the estimated humidity value with the preset humidity threshold, and if the humidity exceeds the humidity threshold, perform dehumidification processing on the power equipment box; S8. During the dehumidification process, continuously monitor the humidity changes at different locations in the power equipment box; S9. During the dehumidification process, detect abnormal conditions of the power equipment cabinet.

2. The automatic dehumidification method in the power equipment box according to claim 1 is characterized in that: The audio signal acquisition in step S1 includes the following steps: Setting environmental noise sensors at different positions in the power equipment box to collect the audio signal, and extracting background noise characteristic parameters and humidity volume characteristics at various positions in the power equipment box through signal processing technology; The preprocessing of the audio signal adopts an adaptive noise reduction algorithm; The preprocessing of the audio signal includes removing noise interference and extracting effective humidity volume feature information.

3. The automatic dehumidification method in the power equipment box according to claim 2 is characterized in that: The step S2 further comprises: After establishing the acoustic propagation model, adjusting the parameters of the adaptive noise reduction algorithm by analyzing the attenuation and distortion degree of the audio signal caused by different materials; The establishment of the acoustic propagation model comprises the following steps: According to the internal structure layout and material information of the power equipment box, the finite element software is used to establish a three-dimensional acoustic scene model, divide the finite element grid, set boundary conditions and load conditions, input material property parameters, and calculate and simulate the acoustic characteristic parameters of different materials; Extract key parameters of sound waves in each material, including propagation velocity, attenuation coefficient and reflection coefficient, and derive calculation results; By analyzing the attenuation and distortion characteristic curves of different materials on audio signals, the influence of each material on the sound signal in different frequency ranges can be determined; Constructing a correspondence between attribute parameters and their acoustic characteristic parameters, wherein the attribute parameters include a material acoustic characteristic database, recording material type, thickness, density and porosity; The acoustic propagation model is established by combining the location of the sound source, the propagation path and the acoustic characteristics of the material.

4. The automatic dehumidification method in the power equipment box according to claim 1 is characterized in that: The step S3 specifically includes: Environmental noise sensors are installed at different locations in the power equipment box to collect noise data samples under different working conditions and form a noise sample database; The signal processing method is used to analyze the collected noise sample data in the time and frequency domain, extract the spectrum characteristics, energy characteristics and statistical characteristics of the noise signal, and construct a noise feature vector data set; Using machine learning algorithms and combining them with noise feature vector datasets, machine learning models are trained to classify and identify different types of noise, obtain noise type labels and intensity feature information, and form a noise identification knowledge base.

5. The automatic dehumidification method in the power equipment box according to claim 1 is characterized in that: The step S4 specifically includes: In combination with the environmental noise characteristics, the acoustic properties of the material and the properties of different positions, the corresponding noise reduction algorithm and parameters are used to perform adaptive noise reduction processing on the audio signal; The adaptive noise reduction processing of the audio signal includes eliminating noise interference and retaining humidity volume feature information.

6. The automatic dehumidification method in the power equipment box according to claim 1 is characterized in that: The step S5 specifically includes: For the audio signal after adaptive noise reduction, a time-frequency analysis method is used to extract volume feature parameters reflecting the humidity state to form a high-dimensional humidity volume feature vector; Perform multi-scale and multi-resolution analysis on the audio signal by wavelet packet decomposition, extract wavelet coefficient statistics at different frequency bands and time scales, and capture non-stationary humidity volume characteristics in the audio signal; The speech feature extraction method is used to obtain the spectrum distribution characteristics reflecting the humidity state, and the dynamic characteristics of volume intensity changing with humidity are obtained; The correlation and significance between each humidity volume feature and humidity state are evaluated through feature selection algorithm; The maximum correlation minimum redundancy algorithm is used to first select the feature with the greatest correlation with the humidity state, and then select the feature with the least correlation with the selected feature from the remaining features until the preset number of features or correlation threshold is met; The chi-square test was used to evaluate whether the feature was independent of the humidity state by calculating the difference between the actual observed frequency and the expected frequency of the feature value.

7. The automatic dehumidification method in the power equipment box according to claim 1 is characterized in that: The step S6 specifically includes: Acquire the audio signals processed by adaptive noise reduction at different positions of different materials in the power equipment box, extract volume feature parameters reflecting the humidity state, construct a multi-dimensional humidity volume feature vector, and form an input data set for the humidity estimation model; A long short-term memory neural network is used to build a deep learning humidity estimation model, and a Dropout layer is added between the LSTM layer and the fully connected layer to prevent the model from overfitting; The mean square error is selected as the loss function of the humidity estimation model, and the Adam optimizer is used to adaptively adjust the model parameters. The model parameters with the lowest mean square error during the training process are saved as the optimal model. The humidity volume feature data of different materials at different locations are divided into training set, validation set and test set by stratified sampling method. According to the distribution of humidity labels, the data set is divided into multiple humidity intervals, and then samples are randomly selected in each interval at a ratio of 8:1:

1. The random partitioning method is used to repeat the K-times data set partitioning, and the model is trained and evaluated each time. Finally, the average of the K-times results is taken as the final performance indicator of the model. The training set is used to train the deep learning humidity estimation model, the validation set is used to evaluate the model performance and tune the parameters, and the test set is used to test the generalization ability of the model; During the training process, the hyperparameters of the LSTM layer are dynamically adjusted to optimize the model's fitting ability and generalization performance. At the same time, an early stopping mechanism is introduced to terminate the training in advance to prevent the model from overfitting. The trained deep learning humidity estimation model is deployed to an embedded device or a cloud server, and the audio signals of different materials at different positions in the power equipment box are received in real time. The humidity volume characteristic parameters are extracted and input into the model for humidity estimation to obtain the real-time humidity value.

8. The automatic dehumidification method in the power equipment box according to claim 7, characterized in that: The step S7 specifically includes: If the humidity estimation value of a certain location in the power equipment box output by the deep learning humidity estimation model is higher than the preset humidity threshold, a humidity abnormality alarm message is automatically generated, the time, location and humidity value of the abnormality are recorded, and the automatic dehumidification device at that location is triggered to start, and dehumidify the corresponding location in the power equipment box until the humidity returns to the theoretical normal humidity value.

9. The automatic dehumidification method in the power equipment box according to claim 8, characterized in that: The step S8 further comprises: According to the humidity changes at different positions in the power equipment box, the working parameters of the automatic dehumidification device are adjusted in real time, and the working parameters include dehumidification intensity and duration.

10. The automatic dehumidification method in the power equipment box according to claim 1, characterized in that: The step S9 further comprises: If no abnormality is detected, the original dehumidification working state is maintained; If an abnormal situation is detected, the dehumidification work will be suspended, the alarm mechanism will be triggered, and an alarm message and data recording the abnormal situation will be sent.

Images