Machine room noise analysis and judgment method and noise analysis and judgment device

Abstract

The invention provides a machine room noise analysis and judgment method and a noise analysis and judgment device, and the method comprises the steps: collecting the original audio of a machine room environment, and carrying out the preprocessing of the collected original audio; extracting a first Mel spectrogram of the preprocessed audio as a time-frequency feature; inputting the extracted first Mel spectrogram into a convolutional neural network model, and training the convolutional neural network model to obtain a trained optimization model; acquiring a second Mel spectrogram of the real-time noise of the machine room, and inputting the second Mel spectrogram into the optimization model; judging whether the steel wire rope is broken or not through the optimization model; if yes, an alarm signal is generated, and alarm information is sent to maintenance personnel. According to the method, the accuracy of noise recognition and judgment of the elevator machine room can be improved, so that whether steel wire strand breakage occurs in the elevator or not can be timely judged, and safe operation of the elevator is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of battery technology, and in particular to a method and device for analyzing and judging noise in a computer room. Background Technology

[0002] In the operation of machine room equipment such as elevators and cranes, broken strands in wire ropes are a significant factor affecting normal equipment operation. For example, in elevators, the wire rope is a critical load-bearing component, playing a vital role in suspending the car and counterweight. When a strand in the wire rope breaks, it directly impacts the safe operation of the elevator: a broken strand reduces the wire rope's load-bearing capacity, potentially failing to stably support the weight of the car and its passengers or cargo, posing a safety hazard of car collapse. Therefore, daily operation requires monitoring of machine room equipment such as elevators and cranes. Noise monitoring devices are typically used to monitor the noise in the elevator machine room to determine if a malfunction has occurred.

[0003] Commonly used noise monitoring methods for computer rooms typically employ sound pressure level threshold alarms. This involves using a microphone to collect ambient noise, triggering an alarm when the decibel level exceeds a preset threshold. However, this method suffers from frequent false alarms because ambient noise (such as conversations or equipment impacts) overlaps with the specific metallic friction sound produced by broken strands in steel wire ropes. Relying solely on volume thresholds fails to effectively distinguish between fault-related and non-fault-related sounds. Furthermore, existing noise monitoring devices cannot differentiate the acoustic characteristics of broken strands in steel wire ropes, damaged bearings, or other mechanical faults, making it difficult to provide accurate fault location information.

[0004] Therefore, it is necessary to improve the existing noise monitoring of elevator machine rooms to overcome the shortcomings of the existing technology. Summary of the Invention

[0005] To overcome the problems existing in related technologies, one of the objectives of this invention is to provide a method for analyzing and judging machine room noise. This method can improve the accuracy of identifying and judging elevator machine room noise, thereby enabling timely detection of whether the elevator has experienced a broken wire strand, which helps to ensure the safe operation of the elevator.

[0006] A method for analyzing and judging noise in a computer room includes:

[0007] Collect raw audio from the computer room environment and preprocess the collected raw audio.

[0008] The first Mel spectrogram of the preprocessed audio is extracted as a time-frequency feature;

[0009] The extracted first Mel spectrogram is input into the convolutional neural network model to train the convolutional neural network model and obtain the trained optimized model.

[0010] Obtain the second Mel spectrum of the real-time noise in the computer room and input the second Mel spectrum into the optimization model;

[0011] The optimized model determines whether a strand of the wire rope has broken; if so, an alarm signal is generated and an alarm message is sent to maintenance personnel.

[0012] In a preferred embodiment of the present invention, the preprocessing of the acquired raw audio includes:

[0013] The original audio is resampled to a uniform sampling rate and then divided into audio frames of fixed duration.

[0014] Noise reduction processing is performed on the audio frames.

[0015] In a preferred embodiment of the present invention, the noise reduction processing of the audio frame includes:

[0016] Extract the first and last 300ms segments from the original audio frame as pure noise segments;

[0017] Calculate the average power spectral density of the noise segment;

[0018] Noise reduction is achieved using spectral subtraction.

[0019] In a preferred embodiment of the present invention, the noise reduction processing of the audio frame includes:

[0020] Noise reduction using deep learning methods includes:

[0021] Acquiring training data: Collect noisy frequency-clean audio pairs, process the acquired data, and obtain the Mel spectrogram of the noisy frequency frame;

[0022] Build and train the model;

[0023] Input the Mel spectrogram of the noisy frequency frame into the training model to obtain the denoised Mel spectrogram.

[0024] In a preferred embodiment of the present invention, extracting the first Mel-spectrum of the preprocessed audio includes:

[0025] The Mel-scale spectrum is calculated using the librosa.feature.melspectrogram function.

[0026] Normalize the spectrogram dimensions to a preset dimension.

[0027] In a preferred embodiment of the present invention, the convolutional neural network model includes:

[0028] Input layer: for the first Mel spectrogram;

[0029] Convolutional and pooling layers: used to extract spatial features from spectrograms;

[0030] Fully connected layer and output layer: Output binary classification probability.

[0031] In a preferred embodiment of the present invention, training the convolutional neural network model includes:

[0032] Use a labeled audio dataset as the model input, where the labels include normal state and broken state;

[0033] The model was trained using a binary cross-entropy loss function and the Adam optimizer.

[0034] A second objective of this invention is to provide a noise analysis and judgment device for implementing the computer room noise analysis and judgment method described above. The device includes:

[0035] Audio acquisition module: Located in the computer room, used to acquire ambient audio;

[0036] Preprocessing module: configured to perform the audio preprocessing steps described above;

[0037] Deep learning recognition module: includes a pre-trained CNN model for classifying Mel spectrograms and outputting the result of the broken strand state;

[0038] IoT communication module: used to send alarm information to remote terminals when a stock breakage is detected.

[0039] The beneficial effects of this invention are as follows:

[0040] This invention provides a method and device for analyzing and judging noise in a computer room. The method includes: acquiring raw audio data of the computer room environment and preprocessing the acquired audio; extracting a first Mel-frequency spectrogram of the preprocessed audio as a time-frequency feature; inputting the extracted first Mel-frequency spectrogram into a convolutional neural network model to train the model and obtain a trained optimized model; acquiring a second Mel-frequency spectrogram of the real-time noise in the computer room and inputting it into the optimized model; determining whether a wire rope has broken strands using the optimized model; if so, generating an alarm signal and sending an alarm message to maintenance personnel. This method preprocesses the acquired noise, improving the accuracy of noise identification during training. By learning the differences in sound characteristics between normal and broken strands in the wire rope and training the model to obtain a trained optimized model, and using this optimized model to monitor real-time noise, the accuracy of broken strand noise identification can be improved, the false alarm rate reduced, and the influence of interference factors such as human speech and environmental noise effectively avoided.

[0041] This application also provides a noise analysis and judgment device including a noise analysis and judgment device for implementing the machine room noise analysis and judgment method described above. The device can be installed in the elevator machine room and used to monitor the elevator's operating status based on the noise in the machine room, determine whether the elevator has experienced a strand breakage, and thus notify maintenance personnel in the early stages of a strand breakage to prevent the potential danger from escalating and ensure elevator safety. Attached Figure Description

[0042] Figure 1 This is a flowchart of the data center noise analysis and judgment method provided in the embodiments of the present invention;

[0043] Figure 2 This is a flowchart of noise reduction processing of audio frames using spectrum provided in an embodiment of the present invention;

[0044] Figure 3 This is a logic diagram of the computer room noise analysis and judgment method provided in the embodiments of the present invention. Detailed Implementation

[0045] Preferred embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0046] In the operation of machine room equipment such as elevators and cranes, broken strands in wire ropes are a significant factor affecting normal equipment operation. For example, in elevators, the wire rope is a critical load-bearing component, playing a vital role in suspending the car and counterweight. When a strand in the wire rope breaks, it directly impacts the safe operation of the elevator: a broken strand reduces the wire rope's load-bearing capacity, potentially failing to stably support the weight of the car and its passengers or cargo, posing a safety hazard of car collapse. Therefore, daily operation requires monitoring of machine room equipment such as elevators and cranes. Noise monitoring devices are typically used to monitor the noise in the elevator machine room to determine if a malfunction has occurred.

[0047] Commonly used noise monitoring methods for computer rooms typically employ sound pressure level threshold alarms. This involves using a microphone to collect ambient noise, triggering an alarm when the decibel level exceeds a preset threshold. However, this method suffers from frequent false alarms because ambient noise (such as conversations or equipment impacts) overlaps with the specific metallic friction sound produced by broken strands in steel wire ropes. Relying solely on volume thresholds fails to effectively distinguish between fault-related and non-fault-related sounds. Furthermore, existing noise monitoring devices cannot differentiate the acoustic characteristics of broken strands in steel wire ropes, damaged bearings, or other mechanical faults, making it difficult to provide accurate fault location information.

[0048] Based on this, this application provides a method for analyzing and judging noise in a computer room.

[0049] Example 1

[0050] like Figures 1-3 As shown in the figure, this embodiment provides a method for analyzing and judging computer room noise, including:

[0051] S100: Collects raw audio from the computer room environment and preprocesses the collected raw audio.

[0052] In this step, the normal operating sounds of 100 elevators of different models under no-load, full-load, and half-load conditions can be continuously collected, accumulating 1000 hours of data, covering environmental background sounds at different times (such as day and night). Twenty sections of steel wire rope with broken strands (ranging from 1 to 3 broken strands) are selected and installed in the test elevators to simulate the sound of broken strands under normal operating conditions, accumulating 200 hours of data collection, including characteristic sounds at the initial, middle, and severe stages of broken strands.

[0053] The preprocessing process includes:

[0054] Data labeling: A combination of manual review and automatic labeling tools is used to add tags to all audio files (0 = normal operation, 1 = disconnected), ensuring a labeling accuracy rate of ≥99%.

[0055] Unified sampling and framing: Using the pydub tool, all audio was resampled to 16kHz, divided into fixed-length segments of 2 seconds each, and invalid segments with more than 80% silence were removed, resulting in 800,000 valid samples (600,000 normal samples and 200,000 broken samples).

[0056] Noise reduction: Spectral subtraction was applied to all samples to remove environmental noise (such as the sound of air conditioning in the computer room and the sound of people walking around), while retaining the core sound characteristics of the wire rope operation.

[0057] S200: Extract the first Mel spectrum of the preprocessed audio as a time-frequency feature;

[0058] The librosa library was used to extract Mel spectrograms from the preprocessed audio segments. The parameters were set as follows: number of Mel filters 128, time frame length 128, resulting in a two-dimensional time-frequency feature map with a shape of 128×128×1, which was used as the model input data.

[0059] S300. Input the extracted first Mel spectrum into the convolutional neural network model, train the convolutional neural network model, and obtain the trained optimized model.

[0060] The training process includes the following steps:

[0061] Model training phase

[0062] Dataset partitioning: The samples were divided into a training set (560,000 samples), a validation set (120,000 samples), and a test set (120,000 samples) in a ratio of 7:1.5:1.5 to ensure that the proportion of normal and broken samples in each set is consistent.

[0063] Model construction: Build a CNN model, which includes 3 convolutional layers (using 32, 64, and 128 3×3 convolutional kernels respectively), 2 pooling layers (2×2 max pooling), 1 fully connected layer (128 neurons) and an output layer (sigmoid activation function, outputting the probability of strand breakage).

[0064] Training configuration: Binary Crossentropy was used as the loss function, Adam optimizer (initial learning rate 3e-4, decaying by 10% every 5 epochs), and iterative training was performed on an NVIDIA Tesla V100 graphics card for 30 epochs. Training was stopped when the F1-Score on the validation set showed no improvement for 5 consecutive epochs.

[0065] S400: Obtain the second Mel spectrum of the real-time noise in the computer room, and input the second Mel spectrum into the optimization model;

[0066] Real-time audio acquisition: The microphone acquires the sound of the computer room in real time, generating an audio segment every 2 seconds. The frame segmentation and noise reduction processing in step 2 are repeated to extract the second Mel spectrogram.

[0067] Elevator failure detection and alarm: The second Mel spectrum is input into the optimization model. When the output failure probability is ≥0.85 (the threshold is determined by the test set), it is determined to be a failure state. An alarm message is sent to the maintenance personnel's mobile APP through the Internet of Things platform (such as MQTT protocol), which includes information such as elevator number, failure probability, and real-time audio clip.

[0068] S500 determines whether the wire rope has broken strands through an optimized model; if so, it generates an alarm signal and sends an alarm message to maintenance personnel.

[0069] The microphone collects the sound from the machine room in real time, generating an audio segment every 2 seconds. The frame segmentation and noise reduction processing in step 2 is repeated to extract the second Mel spectrogram. The second Mel spectrogram is input into the optimization model. When the output failure probability is ≥0.85 (the threshold is verified by the test set), it is determined to be a failure state. An alarm message is sent to the maintenance personnel's mobile APP through an IoT platform (such as MQTT protocol), which includes information such as elevator number, failure probability, and real-time audio segment.

[0070] The aforementioned method for analyzing and judging noise in a computer room, along with its associated device, includes: collecting raw audio data from the computer room environment and preprocessing the collected audio; extracting a first Mel-frequency spectrogram from the preprocessed audio as a time-frequency feature; inputting the extracted first Mel-frequency spectrogram into a convolutional neural network model for training, resulting in a trained optimized model; acquiring a second Mel-frequency spectrogram of the real-time noise in the computer room and inputting it into the optimized model; determining whether a strand of the wire rope has broken through the optimized model; and if so, generating an alarm signal and sending an alarm message to maintenance personnel. This method, by preprocessing the collected noise, improves the accuracy of noise identification during the training process. Furthermore, by learning the differences in sound characteristics between normal and broken strands of the wire rope and training the model to obtain a trained optimized model, and using this optimized model to monitor real-time noise, the accuracy of strand breakage noise identification can be improved, the false alarm rate reduced, and the influence of interference factors such as human speech and environmental noise effectively avoided.

[0071] In a specific implementation, the preprocessing of the acquired raw audio includes:

[0072] The original audio is resampled to a uniform sampling rate and then divided into audio frames of fixed duration.

[0073] Noise reduction processing is performed on the audio frames.

[0074] Furthermore, the noise reduction processing of the audio frames includes:

[0075] S101. Extract the first and last 300ms from the original audio frame as pure noise segments;

[0076] A 300ms segment (containing 13,230 sampling points) was extracted from the beginning and end of each frame of the original audio as a pure noise segment. The selection criteria were as follows: In the computer room audio frame, the first and last 300ms usually do not have the main sound source of the steel wire rope operation, and are mainly environmental noise (such as equipment heat dissipation sound and air flow sound), which can be approximated as pure noise.

[0077] S102. Calculate the average power spectral density of the noise segment;

[0078] Perform Fourier transforms on the first and last pure noise segments to obtain their respective power spectral densities; average the two power spectral densities as the baseline value of the noise power spectral density for that audio frame. The calculation formula is as follows:

[0079]

[0080] Where Pnoise1(f) is the power spectral density of the noise segment at the beginning of the audio frame, Pnoise2(f) is the power spectral density of the noise segment at the end of the frame, and f is the frequency.

[0081] S103. Noise reduction is achieved by using spectral subtraction.

[0082] Perform a Fourier transform on the entire original audio frame to obtain the signal power spectral density Psignal(f); then subtract the average noise power spectral density from the signal power spectrum using spectral subtraction, as shown in the following formula:

[0083] P denoised (f)=max(P signal (f)-P noise (f),∈)

[0084] Here, ∈ represents a minimum value (e.g., 1e-10) to avoid negative power spectra; finally, the processed power spectrum is converted back to the time domain through inverse Fourier transform to obtain the denoised audio frame.

[0085] The average power spectral density is calculated based on the noise segment, making the noise model more consistent with the actual environmental noise of the current audio frame, thus avoiding excessive or insufficient noise reduction due to fluctuations in environmental noise.

[0086] Spectral subtraction specifically removes components that match the noise power spectrum, while retaining the effective audio characteristics of the wire rope operation (such as the high-frequency vibration sound when the strand breaks), providing a high-quality data foundation for subsequent Mel spectrum extraction and model recognition.

[0087] Furthermore, the noise reduction processing of the audio frames includes:

[0088] Noise reduction using deep learning methods includes:

[0089] Acquiring training data: Collect noisy frequency-clean audio pairs, process the acquired data, and obtain the Mel spectrogram of the noisy frequency frame;

[0090] Build and train the model;

[0091] Input the Mel spectrogram of the noisy frequency frame into the training model to obtain the denoised Mel spectrogram.

[0092] Specifically, the implementation process of noise reduction using deep learning is as follows:

[0093] Pure audio recording: In an anechoic laboratory environment, the operating audio of the elevator wire rope during normal operation (without broken strands) and during operation in the state of broken strands (without environmental noise interference) was recorded as "pure audio", with a total duration of 200 hours;

[0094] Environmental noise collection: Record various environmental interference noises (such as conversations, equipment cooling fans, tool collisions, etc.) in the computer room for a total of 150 hours;

[0095] Constructing noisy frequency-clean audio pairs: Using an audio mixing algorithm, clean audio is mixed with environmental noise at different signal-to-noise ratios (5dB-30dB) to generate "noisy frequencies", forming a 1:1 corresponding "noisy frequency-clean audio pair" dataset;

[0096] Audio framing: All audio pairs are divided into 2-second frames (each frame contains 32,000 sampling points), and segments shorter than 2 seconds are removed;

[0097] Mel spectrogram extraction: The librosa library is used to extract Mel spectrograms from the noisy frequency frames and the corresponding clean audio frames respectively (parameters: number of Mel filters 128, number of Fourier transform points 1024, hop length 256), resulting in a noisy spectrogram (input) and a clean spectrogram (target output) with a size of 128×128×1.

[0098] A convolutional neural network based on the U-Net architecture is used as the noise reduction model, with the following structure:

[0099] Input layer: Receive Mel spectrogram (128×128×1) of noisy frequency frames;

[0100] The encoder contains four convolutional blocks, each consisting of two 3×3 convolutional layers (ReLU activation) and one 2×2 max pooling layer. It progressively extracts the feature differences between noise and signal, and outputs 16×16×256 compressed features.

[0101] Decoder: Contains 4 deconvolution blocks, each consisting of 1 2×2 deconvolution layer (upsampling) + 2 3×3 convolution layers (ReLU activation), gradually restoring the feature map size, and finally outputting a 128×128×1 spectrogram;

[0102] Output layer: Outputs a denoised Mel spectrum with the same size as the input through a 1×1 convolutional layer.

[0103] Dataset partitioning: The noisy-clean spectrogram pairs were divided into a training set (68,000 pairs) and a validation set (17,000 pairs) in an 8:2 ratio;

[0104] Training configuration: The loss function is mean squared error (MSE), the optimizer is Adam (initial learning rate 1e-4), the batch size is 32, the training is performed for 30 epochs, and training is stopped when the validation set loss decreases to less than 0.001 for 5 consecutive epochs.

[0105] Model saving: Saves the converged denoising model after training for subsequent real-time denoising.

[0106] Real-time audio acquisition: The audio acquisition module acquires one frame of computer room audio every 2 seconds (including the sound of steel wire rope running and environmental noise);

[0107] Noisy spectrogram extraction: Extract Mel spectrograms from real-time audio frames (parameters consistent with the training phase);

[0108] Deep learning denoising: Input the noisy spectrogram into the pre-trained U-Net denoising model, and the model outputs the denoised Mel spectrogram (removing environmental noise interference and retaining the core acoustic features of the wire rope operation);

[0109] Thread breakage detection: Input the noise-reduced Mel spectrogram into the thread breakage recognition CNN classification model (binary classification, output "normal" or "thread breakage"). If it is determined to be a thread breakage, the alarm module immediately sends an alarm message containing the time and computer room number.

[0110] Verification using a test set (containing 10,000 pairs of noisy-clean audio frames) showed that the signal-to-noise ratio of the audio improved by an average of 18 dB after noise reduction, and the energy proportion of environmental noise in the Mel spectrogram decreased from the original 35%-60% to below 5%; combined with the broken strand identification model, the overall broken strand judgment accuracy reached 99.1%.

[0111] Deep learning models are trained on a large number of noisy-clean audio pairs and can learn the feature patterns of various environmental noises (stationary and non-stationary), such as the speech spectrum distribution of people talking and the transient high-frequency features of tool collisions.

[0112] The U-Net architecture can capture subtle differences between noise and effective signal in the Mel spectrogram. Even in low signal-to-noise ratio (SNR) scenarios (such as 5dB), it can accurately separate ambient noise from the sound of the wire rope running, resulting in higher signal integrity after noise reduction.

[0113] In the noise-reduced Mel spectrum, the proportion of environmental noise energy is significantly reduced (≤5%), and the characteristic spectrum of wire rope operation (such as the energy enhancement in the 2kHz-5kHz band when a strand breaks) is more prominent.

[0114] Specifically, extracting the first Mel spectrogram of the preprocessed audio includes:

[0115] The Mel-scale spectrum is calculated using the librosa.feature.melspectrogram function.

[0116] Normalize the spectrogram dimensions to a preset dimension.

[0117] The specific steps are as follows: Calculate the Mel-scale spectrum using the librosa.feature.melspectrogram function from Python's librosa library, with the following parameter settings:

[0118] Sampling rate (sr): 16kHz (consistent with the uniform sampling rate in the preprocessing stage);

[0119] Number of Mel filter banks (n_mels): 128;

[0120] Fourier transform points (n_fft): 2048;

[0121] hop length (hop_length): 512;

[0122] Power spectral type: 2.0 (Calculate power spectral density).

[0123] Through the above function processing, each audio frame will generate an initial Mel spectrum matrix with dimensions [128, t], where 128 corresponds to the number of Mel filters and t is the number of time frames (determined by the audio frame length and hop length. For a 2-second audio frame, the calculation process for t is: 2 seconds × 16000 sampling points / second = 32000 sampling points, 32000 ÷ 512 = 62.5, after rounding, t = 63, that is, the initial dimension is 128 × 63).

[0124] To ensure uniformity of the dimensions of the Mel spectrograms input to the CNN model, the initial Mel spectrograms need to be normalized to a preset dimension [128, 128]. The specific implementation method is as follows:

[0125] Linear interpolation was used to stretch the time axis (t dimension) of the initial spectrogram, adjusting the 128×63 matrix to 128×128;

[0126] During the interpolation process, the dimension of the Mel filter bank (128) remains unchanged, and only the time frame dimension is expanded to ensure that the proportion of time-frequency features in the spectrogram is not significantly distorted.

[0127] After the above processing, the first Mel spectrogram obtained is fixed at 128×128 in dimension and is a single channel (grayscale image), which can be directly used as the input feature of the CNN model (shape [128,128,1]).

[0128] By standardizing the dimensions, the Mel-scale spectrograms generated from audio frames of different durations (which may have slight differences during frame segmentation) are made dimensional consistently, meeting the input format requirements of CNN models and ensuring the stability of model training and inference. Mel-scale spectrum calculation, compared to ordinary spectrograms, better reflects the characteristics of human hearing, highlighting the energy distribution of wire rope running noise (especially the characteristic acoustic signature when a strand breaks) in specific frequency bands. This provides effective feature data for subsequent models to distinguish between normal and broken strand states.

[0129] More specifically, the convolutional neural network model includes:

[0130] Input layer: for the first Mel spectrogram; the input layer is used to receive the first Mel spectrogram extracted after preprocessing, as the input feature of the model.

[0131] Convolutional and pooling layers: used to extract spatial features from the spectrogram; specifically, convolutional and pooling layers may include:

[0132] Convolutional layer 1:

[0133] Number of convolution kernels: 32, size 3×3;

[0134] Step size: 1;

[0135] Filling method: same (edge ​​filling, making the output feature map size the same as the input);

[0136] Activation function: ReLU, used to introduce non-linear features and enhance the model's ability to capture complex patterns;

[0137] Output feature map size: 128×128×32.

[0138] Pooling layer 1:

[0139] Pooling method: 2×2 max pooling;

[0140] Step size: 2;

[0141] Function: Downsamples the feature map output by convolutional layer 1, preserving key features while reducing the amount of data and avoiding overfitting;

[0142] Output feature map size: 64×64×32.

[0143] Convolutional layer 2:

[0144] Number of convolution kernels: 64, size 3×3;

[0145] Step size: 1;

[0146] Fill method: same;

[0147] Activation function: ReLU;

[0148] Output feature map size: 64×64×64.

[0149] Pooling layer 2:

[0150] Pooling method: 2×2 max pooling;

[0151] Step size: 2;

[0152] Output feature map size: 32×32×64.

[0153] Convolutional layer 3:

[0154] Number of convolution kernels: 128, size 3×3;

[0155] Step size: 1;

[0156] Fill method: same;

[0157] Activation function: ReLU;

[0158] Output feature map size: 32×32×128.

[0159] Pooling layer 3:

[0160] Pooling method: 2×2 max pooling;

[0161] Step size: 2;

[0162] Output feature map size: 16×16×128.

[0163] By alternating operations of multi-layer convolution and pooling, spatial features in the Mel spectrogram are gradually extracted (such as energy distribution patterns in specific frequency bands under broken strand conditions, spectral patterns during normal operation, etc.), progressing from low-level features (edges, textures) to high-level features (discriminative voiceprint patterns), providing a valid basis for subsequent classification.

[0164] Fully connected layer and output layer: Output binary classification probability.

[0165] The fully connected layer and output layer include:

[0166] flatten layer:

[0167] Function: Flatten the 3D feature map (16×16×128) output by pooling layer 3 into a 1D vector with dimensions of 16×16×128=32768, so as to connect it with the fully connected layer.

[0168] Fully connected layer 1:

[0169] Number of neurons: 1024;

[0170] Activation function: ReLU;

[0171] Function: Performs nonlinear transformation on the flattened feature vectors and fuses the extracted features.

[0172] Fully connected layer 2 (output layer):

[0173] Number of neurons: 2 (corresponding to the two categories of "normal state" and "fractured state");

[0174] Activation function: Softmax;

[0175] Function: Outputs binary classification probabilities, namely the probability that the input Mel spectrogram belongs to the "normal state" and the probability that it belongs to the "broken stock state", and the sum of the two probabilities is 1.

[0176] In practical applications, training a convolutional neural network model includes:

[0177] Use a labeled audio dataset as the model input, where the labels include normal state and broken state;

[0178] The model was trained using a binary cross-entropy loss function and the Adam optimizer.

[0179] The model training configuration and process are as follows:

[0180] Loss function: The binary cross-entropy loss function is used, and its calculation formula is as follows:

[0181]

[0182] Where y is the true label of the sample (0 or 1), and y^ is the predicted probability of "broken stock state" output by the model. This function can effectively measure the difference between the predicted value and the true value in a binary classification task, guiding the model to optimize in the direction of reducing the probability of misclassification.

[0183] Optimizer: The Adam optimizer was selected, with an initial learning rate set to 3e-4. The Adam optimizer combines the advantages of momentum algorithms and adaptive learning rates, which can accelerate model convergence and automatically adjust the learning rate in the later stages of training to improve the accuracy of parameter optimization.

[0184] Other parameters:

[0185] Batch size: 32 (32 samples are input for parameter updates in each iteration);

[0186] Training epochs: 50. If the accuracy on the validation set does not improve for 5 consecutive epochs, an early stopping mechanism is triggered to prevent the model from overfitting.

[0187] The training process is as follows:

[0188] Forward propagation: The Mel spectrogram of the training set is input into the convolutional neural network (CNN). The spatial features of the spectrogram are extracted through convolutional and pooling layers. After processing by fully connected layers, the output layer outputs the probability distribution of "normal state" and "broken strand state".

[0189] Loss calculation: The loss value is calculated using the binary cross-entropy loss function based on the predicted probability of the output layer and the true label.

[0190] Backpropagation: Using the Adam optimizer, the weight parameters of each layer in the network (such as convolutional kernel weights and fully connected layer weights) are adjusted backward based on the loss value to minimize the loss function.

[0191] Validation and Adjustment: After each round of training, use the validation set to evaluate the model performance (such as accuracy and F1-Score), and dynamically adjust the learning rate based on the validation results (such as decaying to 1 / 10 of the original rate every 5 rounds) or terminate training.

[0192] After training, the model performance is evaluated using a test set. Key metrics include:

[0193] Accuracy: The proportion of samples correctly classified by the model out of the total number of samples, with a target value of ≥95%;

[0194] F1-Score: A metric that combines precision and recall (especially suitable for imbalanced data scenarios), with a target value of ≥94%.

[0195] Through the above training process, the model can learn the differences in audio features between normal operation and strand breakage of the wire rope, providing reliable model support for strand breakage judgment in subsequent real-time detection.

[0196] Example 2

[0197] like Figures 1-3 As shown in the figure, the specific implementation process of the above-mentioned computer room noise analysis and judgment method in this embodiment is detailed as follows:

[0198] (1) Raw audio acquisition and preprocessing

[0199] Audio Acquisition: Control 4 microphones to continuously collect audio data from the elevator machine room, including scenarios such as normal elevator operation (no broken strands), operation with broken strands in the steel wire rope, and environmental interference (such as staff conversations and equipment cooling fan noise), with a total collection time of 1000 hours, forming a raw audio dataset.

[0200] Preprocessing:

[0201] Unified sampling rate: All original audio files were resampled to 16kHz using the librosa library to eliminate the impact of sampling rate differences on feature extraction;

[0202] Frame segmentation: The audio is divided into fixed-length segments of 2 seconds each (each segment contains 32,000 sampling points), and segments shorter than 2 seconds are discarded;

[0203] Noise reduction: Spectral subtraction is used to reduce noise in each audio segment. First, the silent segment is extracted from the audio clip as a noise template. Then, the audio is converted to the frequency domain through Fourier transform, and the spectral energy of the noise template is subtracted. Finally, the audio is converted back to the time domain through inverse Fourier transform to remove environmental interference noise.

[0204] (2) Time-frequency feature extraction

[0205] Extract the Mel spectrogram of the preprocessed audio using the librosa.feature.melspectrogram function from the librosa library:

[0206] Parameter settings: Mel filter bank number 128, Fourier transform points 2048, hop length 512;

[0207] Output: Each 2-second audio segment corresponds to a 128×128×1 Mel spectrogram (height 128 represents the number of Mel filters, width 128 represents the number of time frames, and channel 1 is a single-channel grayscale image), which serves as the input feature of the CNN model (i.e., the first Mel spectrogram).

[0208] (3) CNN Model Construction and Training

[0209] Model structure:

[0210] Input layer: Receives a 128×128×1 Mel spectrogram;

[0211] Convolutional layer 1: 32 3×3 convolutional kernels, stride 1, activation function ReLU, output feature map size 126×126×32;

[0212] Pooling layer 1: 2×2 max pooling, stride 2, output 63×63×32;

[0213] Convolutional layer 2: 64 3×3 convolutional kernels, stride 1, ReLU activation function, output 61×61×64;

[0214] Pooling layer 2: 2×2 max pooling, step size 2, output 30×30×64;

[0215] Fully connected layer 1: 1024 neurons, ReLU activation function;

[0216] Fully connected layer 2: 2 neurons (corresponding to the "normal" and "fractured" categories), activation function Softmax;

[0217] Training configuration:

[0218] Dataset partitioning: The labeled Mel spectrograms were divided into a training set (700 hours of data), a validation set (150 hours of data), and a test set (150 hours of data) in a ratio of 7:1.5:1.5. The label "0" represents normal operating audio and the label "1" represents operating audio with broken strands.

[0219] Loss function: Binary Crosssentropy is used;

[0220] Optimizer: Adam optimizer, initial learning rate 3e-4, decaying by 10% every 5 epochs;

[0221] Training parameters: batchsize=32, training for 50 epochs, training stops when the validation set accuracy does not improve for 5 consecutive epochs;

[0222] Output: The optimized model after training achieves a test set accuracy of 98.6% and an F1-Score of 98.2%.

[0223] (4) Real-time noise detection and alarm

[0224] Real-time data processing: The IoT terminal collects the real-time noise of the computer room every 2 seconds, repeats the preprocessing and feature extraction steps (1)-(2), and obtains the second Mel spectrum.

[0225] Stock breakage judgment: Input the second Mel spectrum into the optimization model, and the model outputs the probability value of the "stock breakage" category;

[0226] Alarm Trigger: If the probability value is ≥0.95 (preset threshold), it is determined that the wire rope strand is broken. The real-time detection and alarm module generates an alarm signal (including the location of the machine room, detection time, and probability of strand breakage) and sends the alarm information to the maintenance personnel's mobile APP and IoT platform through the 4G communication module.

[0227] Example 3

[0228] like Figures 1-3 As shown, this embodiment provides a noise analysis and judgment device for implementing the computer room noise analysis and judgment method described above. The device includes:

[0229] Audio acquisition module: Located in the computer room, used to acquire ambient audio;

[0230] Preprocessing module: configured to perform the audio preprocessing steps as described above; to perform the audio preprocessing steps.

[0231] The preprocessing module verifies the sampling rate of the digital audio data transmitted by the audio acquisition module to ensure that all audio data is uniformly 16kHz (if there is a difference, resampling is performed); the audio data is divided into frames according to a fixed duration (e.g., 2 seconds / segment), with each audio segment as an independent processing unit; a noise reduction algorithm (such as spectral subtraction or a deep learning-based noise reduction model) is used to denoise the framed audio segments, removing irrelevant interference noise from the environment and retaining effective sound features related to the operation of the wire rope; the integrity of the denoised audio segments is checked, and invalid or broken audio frames are removed to ensure the validity of the data for subsequent processing.

[0232] Deep learning recognition module: includes a pre-trained CNN model for classifying Mel spectrograms and outputting the result of the broken strand state;

[0233] The deep learning recognition module incorporates a pre-trained CNN model, which is trained on a large dataset of labeled audio data (including audio data in both normal and broken states). The specific workflow is as follows:

[0234] The system receives the denoised audio segment output by the preprocessing module, calls the librosa.feature.melspectrogram function in the librosa library to extract the Mel spectrogram of the audio segment, and normalizes its size to a preset dimension of 128×128×1.

[0235] The normalized Mel spectrogram is input into a pre-trained CNN model. The model extracts spatial features from the spectrogram through convolutional and pooling layers, and then processes them through a fully connected layer to output the classification result corresponding to the audio segment—that is, the binary classification probability of whether the wire rope is in a normal state or a broken strand state.

[0236] When the probability of the stock breakage state exceeds a preset threshold (e.g., 0.95), it is determined that a stock breakage state has been identified.

[0237] IoT communication module: used to send alarm information to remote terminals when a stock breakage is detected.

[0238] Integrated with either NB-IoT or 4G communication modules (the appropriate communication method is selected based on the data center network environment), this module features low power consumption and stable transmission. It establishes a data connection with the deep learning recognition module. When the deep learning recognition module detects a broken stock condition, it immediately receives the result and related information (such as detection time and data center number), packages this information into an alarm signal, and sends it via the IoT network to a remote terminal (such as a maintenance personnel's mobile app or the data center management platform) for real-time alarm notification.

[0239] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings. In the description of this application, it should be understood that the orientation or positional relationship indicated by directional terms such as "front, back, up, down, left, right", "horizontal, vertical, horizontal" and "top, bottom" is usually based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing this application and simplifying the description. Unless otherwise stated, these directional terms do not indicate or imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the scope of protection of this application; the directional terms "inner" and "outer" refer to the inner and outer contours relative to the outline of each component itself.

[0240] For ease of description, spatial relative terms such as "above," "on top of," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "on top of" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatial relative descriptions used herein will be interpreted accordingly.

[0241] Furthermore, it should be noted that the use of terms such as "first" and "second" to define components is merely for the purpose of distinguishing the corresponding components. Unless otherwise stated, these terms have no special meaning and therefore should not be construed as limiting the scope of protection of this application. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for analyzing and judging noise in a computer room, characterized in that, include: Collect raw audio from the computer room environment and preprocess the collected raw audio. The first Mel spectrogram of the preprocessed audio is extracted as a time-frequency feature; The extracted first Mel spectrogram is input into the convolutional neural network model to train the convolutional neural network model and obtain the trained optimized model. Obtain the second Mel spectrum of the real-time noise in the computer room and input the second Mel spectrum into the optimization model; The optimized model determines whether a strand of the wire rope has broken; if so, an alarm signal is generated and an alarm message is sent to maintenance personnel.

2. The method for analyzing and judging computer room noise according to claim 1, characterized in that: The preprocessing of the acquired raw audio includes: The original audio is resampled to a uniform sampling rate and then divided into audio frames of fixed duration. Noise reduction processing is performed on the audio frames.

3. The method for analyzing and judging computer room noise according to claim 2, characterized in that: The noise reduction process for the audio frames includes: Extract the first and last 300ms segments from the original audio frame as pure noise segments; Calculate the average power spectral density of the noise segment; Noise reduction is achieved using spectral subtraction.

4. The method for analyzing and judging computer room noise according to claim 2, characterized in that: The noise reduction process for the audio frames includes: Noise reduction using deep learning methods includes: Acquiring training data: Collect noisy frequency-clean audio pairs, process the acquired data, and obtain the Mel spectrogram of the noisy frequency frame; Build and train the model; Input the Mel spectrogram of the noisy frequency frame into the training model to obtain the denoised Mel spectrogram.

5. The method for analyzing and judging computer room noise according to any one of claims 1-4, characterized in that: The first Mel spectrogram of the preprocessed audio is extracted, including: The Mel-scale spectrum is calculated using the librosa.feature.melspectrogram function. Normalize the spectrogram dimensions to a preset dimension.

6. The method for analyzing and judging computer room noise according to any one of claims 1-4, characterized in that: The convolutional neural network model includes: Input layer: for the first Mel spectrogram; Convolutional and pooling layers: used to extract spatial features from spectrograms; Fully connected layer and output layer: Output binary classification probability.

7. The method for analyzing and judging computer room noise according to claim 6, characterized in that: Training a convolutional neural network model includes: Use a labeled audio dataset as the model input, where the labels include normal state and broken state; The model was trained using a binary cross-entropy loss function and the Adam optimizer.

8. A noise analysis and judgment device, characterized in that: For implementing the computer room noise analysis and judgment method as described in any one of claims 1-7, the apparatus comprises: Audio acquisition module: Located in the computer room, used to acquire ambient audio; Preprocessing module: configured to perform the audio preprocessing steps as described in any one of claims 1-7; Deep learning recognition module: includes a pre-trained CNN model for classifying Mel spectrograms and outputting the result of the broken strand state; IoT communication module: used to send alarm information to remote terminals when a stock breakage is detected.

Images