Activated carbon adsorption facility maintenance method and system based on life prediction

Abstract

This invention discloses a maintenance method and system for activated carbon adsorption facilities based on lifetime prediction, belonging to the field of industrial equipment condition monitoring technology. The method includes: using an acoustic sensor array deployed in the activated carbon adsorption facility to emit high-frequency sound waves and receive response signals to acquire phonon spectrum data reflecting the lattice vibration of activated carbon; extracting multi-dimensional feature parameters from this data, and constructing a dynamically evolving phonon metabolic fingerprint through dimensionality reduction; performing decay analysis by combining the phonon metabolic fingerprint and real-time operating data to output a predicted value of the remaining lifetime; and finally implementing preventive maintenance based on the predicted value to achieve precise operation and maintenance of the activated carbon adsorption facility. This invention solves the technical problems of traditional maintenance methods relying on macroscopic parameters, low prediction accuracy, and inability to anticipate sudden failures, achieving the technical effect of accurate lifetime prediction and early warning of sudden risks for the operating conditions of activated carbon adsorption facilities.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring technology, and in particular to a method and system for maintaining activated carbon adsorption facilities based on lifespan prediction. Background Technology

[0002] Activated carbon adsorption facilities are widely used in various spaces, and their accurate maintenance is crucial for environmental compliance and long-term equipment operation, with lifespan prediction being the core support. Existing maintenance technologies mostly rely on macroscopic data such as pressure difference and pollutant concentration, and data processing only stays at the level of surface statistics, failing to delve into information related to changes in the microstructure of activated carbon. Affected by fluctuations in operating conditions and the adsorption-desorption cycle, traditional data processing methods cannot accurately correlate microscopic decay with lifespan reduction, resulting in large prediction biases, delayed maintenance, and difficulty in meeting preventive maintenance needs. Summary of the Invention

[0003] This application provides a maintenance method and system for activated carbon adsorption facilities based on lifetime prediction, which solves the technical problems of traditional maintenance methods relying on macroscopic parameters, low prediction accuracy, and inability to respond to sudden failures in advance.

[0004] The first aspect of this application provides a maintenance method for activated carbon adsorption facilities based on lifetime prediction. The method includes: transmitting high-frequency acoustic signals to the activated carbon adsorption facility via an acoustic sensor array deployed thereon, and receiving response signals to obtain phonon spectrum data reflecting the lattice vibration state of the activated carbon; extracting characteristic parameters characterizing changes in the microstructure state of the activated carbon from the phonon spectrum data to construct a dynamically evolving phonon metabolic fingerprint, wherein the phonon metabolic fingerprint is a low-dimensional feature vector obtained by dimensionality reduction after multi-dimensional analysis of the phonon spectrum data; combining the phonon metabolic fingerprint with real-time operating data of the activated carbon adsorption facility to perform activated carbon decay analysis and output a predicted remaining lifetime value when the activated carbon reaches a preset failure threshold; and performing preventative maintenance based on the predicted remaining lifetime value.

[0005] A second aspect of this application provides a maintenance system for activated carbon adsorption facilities based on lifetime prediction. The system includes: a phonon spectrum data acquisition module, used to transmit high-frequency acoustic signals to the activated carbon adsorption facility via an acoustic sensor array deployed thereon, and receive response signals to obtain phonon spectrum data reflecting the lattice vibration state of the activated carbon; a phonon metabolic fingerprint construction module, used to extract characteristic parameters characterizing changes in the microstructure state of the activated carbon from the phonon spectrum data, constructing a dynamically evolving phonon metabolic fingerprint, wherein the phonon metabolic fingerprint is a low-dimensional feature vector obtained by dimensionality reduction after multi-dimensional analysis of the phonon spectrum data; a remaining lifetime prediction value acquisition module, used to combine the phonon metabolic fingerprint and real-time operating condition data of the activated carbon adsorption facility to perform activated carbon decay analysis and output a remaining lifetime prediction value when the activated carbon reaches a preset failure threshold; and a preventative maintenance execution module, used to perform preventative maintenance based on the remaining lifetime prediction value.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application collects relevant data reflecting the internal vibration state of the material by deploying a sensor array within the activated carbon adsorption facility. After processing such as multi-dimensional feature extraction and dimensionality reduction mapping, dynamic evolution feature indicators are constructed. Combined with real-time operating condition data, decay analysis and dynamic switching of attenuation templates are performed. At the same time, the deterioration of local conditions is monitored to identify the risk of sudden failure, thereby accurately predicting the remaining service life of the material and triggering corresponding maintenance. This makes the maintenance of activated carbon adsorption facilities more preventive, targeted, and reliable, achieving the technical effect of accurately predicting the service life of activated carbon adsorption facilities and providing early warning of sudden risks. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a schematic flowchart of the activated carbon adsorption facility maintenance method based on lifespan prediction provided in the embodiments of this application.

[0009] Figure 2 This is a schematic diagram of the structure of the activated carbon adsorption facility maintenance system based on lifespan prediction provided in the embodiments of this application.

[0010] Figure labeling: 1. Phonon spectrum data acquisition module; 2. Phonon metabolic fingerprint construction module; 3. Remaining lifetime prediction value acquisition module; 4. Preventive maintenance execution module. Detailed Implementation

[0011] This application provides a maintenance method and system for activated carbon adsorption facilities based on lifetime prediction, which solves the technical problems of traditional maintenance methods relying on macroscopic parameters, low prediction accuracy, and inability to respond to sudden failures in advance.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, a maintenance method for activated carbon adsorption facilities based on lifespan prediction is provided, wherein the method includes: By using an acoustic sensor array deployed in the activated carbon adsorption facility, high-frequency acoustic signals are emitted into the activated carbon adsorption facility, and response signals are received to obtain phonon spectrum data reflecting the lattice vibration state of activated carbon.

[0015] In this embodiment, activated carbon adsorption facilities are commonly used in the industrial environmental protection field. They are devices or systems that use activated carbon as the core adsorption material to treat pollutants in waste gas, wastewater, etc., to achieve purification, such as adsorption towers and adsorption boxes. The lattice vibration state of activated carbon refers to the thermal vibration or vibration caused by external forces that occurs when atoms and molecules inside the activated carbon crystal revolve around their equilibrium positions. Its state changes directly reflect the integrity of the activated carbon's microstructure and its porosity characteristics.

[0016] Specifically, an acoustic sensor array consisting of multiple microelectromechanical system (MEMS) sensor nodes is used. The sensor nodes are evenly distributed and embedded inside the activated carbon bed. By sending high-frequency acoustic signals at a preset frequency and receiving response signals, phonon spectrum data that can reflect the lattice vibration state of activated carbon is finally obtained.

[0017] Characteristic parameters representing changes in the microstructure of activated carbon are extracted from the phonon spectrum data to construct a dynamically evolving phonon metabolic fingerprint. The phonon metabolic fingerprint is a low-dimensional feature vector obtained by dimensionality reduction after multi-dimensional analysis of the phonon spectrum data.

[0018] Optionally, frequency domain features, time domain features, and spatial correlation features are first extracted from the phonon spectrum data to form an initial high-dimensional feature pool. Then, an autoencoder is used to perform dimensionality reduction and feature selection on the initial high-dimensional feature pool to obtain a core feature vector. Finally, the core feature vector is mapped to a preset performance degradation scale to synthesize a phonon metabolic fingerprint.

[0019] By combining the phonon metabolic fingerprint and the real-time operating data of the activated carbon adsorption facility, the decay analysis of the activated carbon is performed, and the predicted remaining service life of the activated carbon when it reaches the preset failure threshold is output.

[0020] In one embodiment of this application, decay analysis is carried out by combining phonon metabolic fingerprint and real-time operating data of activated carbon adsorption facility. First, the operating mode is identified by analyzing the adsorption impact per unit time corresponding to the operating data. Then, the capacity decay template characterizing repeated adsorption-desorption of activated carbon is dynamically switched according to the operating mode. Finally, the template is used to predict the phonon metabolic fingerprint and output the predicted value of the remaining effective capacity under the corresponding operating mode and the predicted value of the remaining service life of activated carbon when it reaches the preset failure threshold.

[0021] Preventive maintenance is performed based on the predicted remaining useful life.

[0022] Specifically, based on the predicted remaining service life of activated carbon when it reaches the preset failure threshold, its total design life is obtained through actual measurement or manufacturer calibration and divided into graded intervals: The remaining service life ≥ 50% of the total design life is the long-term interval, maintaining a routine operation monitoring frequency of once every 1-2 hours, collecting phonon metabolic fingerprints and real-time operating data according to established rules, and simultaneously recording data fluctuation trends; the remaining service life between 20% and 50% of the total design life is the medium-term interval, adjusting the inlet pollutant concentration to a suitable range through a pre-treatment device, optimizing parameters such as bed wind speed and operating humidity, with the parameter adjustment range determined based on the changing trends of operating conditions and phonon metabolic fingerprints, to slow down the activated carbon decay rate; the remaining service life ≤ 20% of the total design life is the interval near the preset threshold, planning the activated carbon regeneration process or new carbon procurement and reserve plan 3-7 days in advance, continuously tracking phonon metabolic fingerprints and real-time operating data, and stopping the machine at the appropriate time to perform regeneration or replacement operations. If the predicted deviation exceeds ±10% during maintenance, the capacity decay template is immediately recalibrated and the maintenance timing is adjusted to avoid adsorption failure leading to excessive pollutant emissions or facility malfunctions, ensuring the continuous, stable, and efficient operation of the adsorption facility.

[0023] Furthermore, the method provided in this application embodiment includes: The acoustic sensor array is composed of multiple microelectromechanical system (MEMS) sensor nodes uniformly embedded inside the activated carbon bed of the activated carbon adsorption facility; the multiple MEMS sensor nodes transmit sound waves based on a preset frequency and receive response signals to obtain the phonon spectrum data.

[0024] Specifically, firstly, suitable microelectromechanical system (MEMS) sensor nodes are selected. These sensors are characterized by their small size, high sensitivity, and ability to stably transmit and receive high-frequency signals, meeting the monitoring requirements inside the activated carbon bed. Next, the sensor array is evenly embedded. The length, width, and height of the activated carbon bed are measured, and the area is divided into a grid pattern to ensure effective coverage of each area and avoid blind spots. Based on the grid, embedding holes matching the sensor node size are pre-set at the center of each grid. After placing the sensor node into the hole, the gaps are filled with a sealing material compatible with the activated carbon bed material. This both secures the sensor and prevents airflow leakage, ensuring stable sensor operation within the bed.

[0025] Next, high-frequency acoustic signal transmission and response signal reception are performed. First, based on the commonly used frequency range for detecting the microstructure of materials in acoustic testing technology, a transmission frequency of 0.1THz to 2THz is set. This frequency can effectively excite the lattice vibration of activated carbon and form detectable phonon signals. The signal generator module is connected to the microelectromechanical system (MEMS) sensor node. The signal generator generates a high-frequency electrical signal corresponding to the preset frequency. The transmitting unit of the sensor node converts the electrical signal into a high-frequency acoustic signal and transmits it into the activated carbon bed. The output power of the signal generator is calibrated to ensure sufficient acoustic intensity without damaging the activated carbon structure.

[0026] Subsequently, the receiving unit of the microelectromechanical system (MEMS) sensor node captures the reflected and transmitted acoustic response signals within the bed in real time, converts them into electrical signals, and transmits them to the data acquisition module via anti-interference shielded cables. The data acquisition module uses a commonly used high-speed data acquisition card in industry, with a sampling frequency set to more than 10 times the transmission frequency to ensure complete capture of the signal waveform. At the same time, a low-pass filtering method is used to filter out noise from the environment and electromagnetic interference, improving signal purity.

[0027] Finally, the filtered electrical signal is transmitted to the data processing unit, where the time-domain signal is converted into a frequency-domain signal using the Fourier transform method. Characteristic parameters such as frequency and amplitude of the signal are extracted to generate phonon spectrum data that reflects the vibration state of the activated carbon lattice. The Fourier transform is a signal frequency domain conversion method well known to those skilled in the art and can be implemented through professional signal processing software or hardware modules.

[0028] By employing techniques such as microelectromechanical system (MEMS) sensor arrangement, high-frequency signal transmission and reception, signal filtering, and Fourier transform, accurate acquisition of phonon spectrum data inside the activated carbon bed was achieved, providing reliable data support for subsequent extraction of microstructural characteristic parameters of activated carbon.

[0029] Furthermore, the method provided in this application embodiment includes: Frequency domain features, time domain features, and spatial correlation features are extracted from the phonon spectrum data to form an initial high-dimensional feature pool; an autoencoder is used to reduce the dimensionality and select features from the initial high-dimensional feature pool to obtain a core feature vector; the core feature vector is mapped to a preset performance attenuation scale to synthesize the phonon metabolic fingerprint.

[0030] In this embodiment of the application, the autoencoder is an unsupervised deep learning model composed of an encoder and a decoder, used for data processing scenarios such as dimensionality reduction of high-dimensional data, key feature extraction, and data reconstruction.

[0031] Optionally, the acquired raw phonon spectrum data is first preprocessed using low-pass filtering. The cutoff frequency is set to 2.5THz, taking into account the preset frequency range of 0.1THz to 2THz for the high-frequency acoustic signals emitted by the sensor nodes. This cutoff frequency filters out high-frequency noise above 2.5THz, such as environmental electromagnetic interference and sensor noise, accurately preserving the effective information related to activated carbon lattice vibrations in the phonon spectrum signal. Simultaneously, maximum-minimum normalization is used to map the signal amplitudes acquired by different sensor nodes to the [0,1] interval. The mapping formula is y=(x-min(x)) / (max(x)-min(x)), where x is the original signal amplitude, min(x) is the minimum signal amplitude acquired by the sensor node, and max(x) is the maximum signal amplitude acquired by the sensor node. This mapping eliminates the influence of signal amplitude differences between different sensor nodes, ensuring consistency in subsequent feature extraction.

[0032] Subsequently, frequency domain feature extraction, time domain feature extraction, and spatial correlation feature extraction are performed, which will be explained in detail later. The extracted frequency domain feature set, time domain feature set, and spatial correlation feature set are concatenated and integrated according to a unified data format, which is set to 64-bit floating point. Each feature data includes three basic fields: feature name, acquisition timestamp, and corresponding sensor node number. Invalid features are eliminated by a traversal filtering method. The judgment criteria are: using 10 consecutive acquisition cycles as a time window, if the fluctuation range of all values ​​of a feature within the window is less than 0.001, it is judged as an invalid feature with constant value and no change, and is eliminated; all valid features with fluctuation range not less than 0.001 and capable of characterizing the changes in the microstructure state of activated carbon are retained, and finally, a high-dimensional initial feature pool is formed.

[0033] If data is missing, when data from a single sensor node is lost, it is filled by the average of the corresponding features from three adjacent sensor nodes in the same area; when a single feature value is missing, it is filled by the median of that feature in the historical data of the same period (the last 30 acquisition cycles), ensuring the integrity of the feature data. During the integration process, a data matrix concatenation method is used, arranging the frequency domain feature set, time domain feature set, and spatial correlation feature set sequentially as column vectors, combining them into a complete feature matrix. The rows of the matrix correspond to each acquisition time, and the columns correspond to individual features, achieving structured storage of the feature data for easy retrieval during subsequent autoencoder processing.

[0034] Next, the core feature vector is obtained using an autoencoder: First, an autoencoder neural network containing an encoder and a decoder is constructed. An initial high-dimensional feature pool from historical data is used as training samples, and unsupervised training is performed with the goal of minimizing the reconstruction error between the input and the decoder output. During training, the encoder compresses the initial high-dimensional feature pool to a bottleneck layer through multiple nonlinear transformations to obtain the latent representation vector. Furthermore, a sparsity constraint on the activation values ​​of the bottleneck layer is added to the loss function to activate key neurons strongly correlated with the degradation of adsorption performance and suppress redundant features, ultimately completing the reconstruction of the core feature vector. This step will be explained in detail later.

[0035] Finally, the core feature vector is input into a preset performance degradation scale, and the output scalar value between 0 and 1 is used as the phonon metabolic fingerprint, where 0 represents the initial new state of activated carbon and 1 represents the completely failed state that has reached the preset failure threshold. This step will also be explained in detail in the following content.

[0036] Furthermore, the method provided in this application embodiment includes: Frequency domain features include phonon density of states peak parameters, spectral moment characteristics, and resonant trajectory; time domain features include phonon lifetime distribution and waveform distortion; spatial correlation features include heterogeneity index and correlation length.

[0037] Specifically, after performing a Fast Fourier Transform on the preprocessed phonon spectrum time-domain signal, the vibration signal in the time dimension is converted into a frequency-amplitude distribution frequency-domain signal. Frequency domain feature extraction is then performed based on this frequency-domain signal. The specific steps are as follows: When extracting the peak parameters of the phonon density of states, the frequency and amplitude axes of the frequency domain signal are first traversed to locate the point with the largest amplitude, which is the main peak. The frequency value corresponding to this point is the position of the main peak, reflecting the dominant frequency of the activated carbon lattice vibration. When the activated carbon microstructure is damaged or the pores are blocked, the position of the main peak will shift. A horizontal straight line is drawn at the peak amplitude of the main peak, intersecting the main peak curve to obtain two intersection points. The frequency difference between the two intersection points is the full width at half maximum (FWHM). The FWHM reflects the energy concentration of the lattice vibration mode; the FWHM increases when the structural integrity decreases. The areas of the left and right curves of the main peak within the FWHM are calculated. The ratio of the left area to the right area is the asymmetry. The asymmetry reflects the uneven distribution of lattice vibration energy and indicates local anomalies in the microstructure. Divide the frequency domain signal into low-frequency and high-frequency regions, and calculate the area enclosed by the main peak curve and the frequency axis in each region. The ratio of the area of ​​the low-frequency region to the area of ​​the high-frequency region is the peak area ratio. This ratio can reflect the distribution relationship of vibration energy in different frequency bands. Structural decay will cause the ratio to change regularly.

[0038] When extracting spectral moment features, each order of spectral moment is calculated based on the frequency domain signal. The first-order moment is the weighted mean of frequencies, with the weight being the amplitude of the corresponding frequency, reflecting the center of the frequency distribution of the frequency domain signal. The second-order moment is the variance, calculated by the mean of the sum of squares of the deviations of each frequency from the first-order moment, reflecting the degree of dispersion of the frequency distribution. The third-order moment is the skewness, calculated by the mean of the sum of cubes of the deviations of each frequency from the first-order moment, reflecting the direction of asymmetry in the frequency distribution. The fourth-order moment is the kurtosis, calculated by the ratio of the mean of the sum of the fourth powers of the deviations of each frequency from the first-order moment to the square of the variance, reflecting the sharpness of the frequency distribution. The spectral moments of each order together constitute the spectral moment features, comprehensively describing the distribution characteristics of the frequency domain signal and providing multi-dimensional basis for judging the state of microstructure.

[0039] When extracting the resonance peak trajectory, phonon spectrum data is continuously collected at fixed time intervals. The main peak position is extracted for the frequency domain signal at each moment. The main peak positions at all moments are arranged in chronological order to form a continuous curve, which is the resonance peak trajectory. This trajectory can dynamically track the evolution of the dominant frequency of lattice vibration and intuitively reflect the time process of the decay of the microstructure of activated carbon.

[0040] Next, time-domain feature extraction is performed directly based on the preprocessed phonon spectrum time-domain signal. When extracting the phonon lifetime distribution, the decay stage of the time-domain signal is subjected to exponential fitting to obtain the vibrational energy decay curve. The time corresponding to the energy decreasing from the initial amplitude to 1 / e of the initial amplitude in the decay curve is the relaxation time. The relaxation time is calculated for the time-domain signal corresponding to different frequencies. The set of all relaxation times constitutes the phonon lifetime distribution, which reflects the dissipation rate of phonon vibrational energy at different frequencies. When the activated carbon microstructure is intact, the relaxation time is longer and the distribution is concentrated. When the structure is damaged or adsorption is saturated, the relaxation time will be shortened and the distribution will be dispersed.

[0041] When extracting waveform distortion, the phonon spectrum time-domain signal of activated carbon in its virgin state is first acquired as the initial reference signal. The cross-correlation value between the received signal to be detected and the initial reference signal is calculated. The cross-correlation value is divided by the product of the root mean square of the received signal amplitude and the root mean square of the initial reference signal amplitude to obtain the normalized cross-correlation coefficient. This coefficient is the waveform distortion. The waveform distortion reflects the degree of deviation between the received signal and the reference signal in the ideal state. When the pores of the activated carbon bed are blocked or the lattice is distorted, the sound wave propagation path changes, which will lead to a decrease in waveform distortion.

[0042] Then, the phonon spectrum data collected by all uniformly embedded sensor nodes are synchronized using timestamps to ensure that the signals from different nodes are fully aligned in the time dimension. Spatial correlation feature extraction is then performed based on the synchronized multi-node data. When extracting the heterogeneity index, acoustic parameters are first extracted from the phonon spectrum data of each sensor node. The acoustic parameters can be selected as sound velocity or attenuation coefficient. The standard deviation of the acoustic parameters corresponding to all sensor nodes is calculated. The standard deviation is the arithmetic square root of the sum of squares of the deviations of each acoustic parameter from the mean of the parameter, or the entropy value of the acoustic parameter is calculated. The entropy value is obtained by calculating the distribution probability of the acoustic parameter. The heterogeneity index reflects the spatial uniformity of the acoustic parameters inside the activated carbon bed. When the microstructure deteriorates in a local area of ​​the bed, the acoustic parameters of the sensor nodes in that area differ more from those in other areas, leading to an increase in the heterogeneity index.

[0043] When extracting the correlation length, the spatial autocorrelation function of the acoustic parameters between any two sensor nodes is calculated. A curve is plotted with the distance between the two sensor nodes as the x-axis and the autocorrelation function value as the y-axis. The x-axis value corresponding to the decay of the autocorrelation function value to 1 / e in the curve is found. This value is the correlation length. The correlation length reflects the spatial spread range of the microstructure changes of activated carbon. When the microstructure decay spreads from the local to the surrounding area, the correlation length will change accordingly, which can intuitively reflect the spatial propagation trend of decay.

[0044] Furthermore, the method provided in this application embodiment includes: An autoencoder neural network comprising an encoder and a decoder is constructed. An initial high-dimensional feature pool from historical data is used as training samples. The goal is to minimize the reconstruction error between the input features and the decoder output features. The autoencoder neural network is trained unsupervised. The encoder compresses the initial high-dimensional feature pool into a bottleneck layer through multiple nonlinear transformations. The output of neurons in the bottleneck layer is a latent representation vector that integrates the original multi-feature information. During training, sparsity constraints on the activation values ​​of the bottleneck layer are added to the loss function, causing the autoencoder neural network to activate key neurons strongly correlated with the decay of adsorption performance. This automatically suppresses redundant features in the latent representation vector, completing the reconstruction of the core feature vector.

[0045] Specifically, an autoencoder neural network is first constructed, consisting of an encoder and a decoder connected in series, with signal transmission between them achieved through a bottleneck layer. The encoder structure consists of an input layer, three hidden layers, and a bottleneck layer. The number of neurons in the input layer is consistent with the feature dimension of the initial high-dimensional feature pool to ensure complete reception of high-dimensional feature data. The number of neurons in the first hidden layer is set to half that of the input layer, the second hidden layer to half that of the first layer, the third hidden layer to half that of the second layer, and the bottleneck layer to 1 / 10 that of the input layer. All layers are connected using a fully connected method, meaning that each neuron in the previous layer establishes a weighted connection with all neurons in the next layer.

[0046] The encoder employs the ReLU activation function in each hidden layer, introducing a non-linear transformation through the operation f(x) = max(0,x) to avoid the gradient vanishing problem. The bottleneck layer does not have an activation function and directly outputs the latent representation vector. The decoder structure is symmetrical to the encoder, consisting of a bottleneck layer, three hidden layers, and an output layer. The number of neurons in each layer increases inversely to the corresponding layer in the encoder; that is, the number of neurons in the first hidden layer of the decoder is the same as that in the third hidden layer of the encoder, increasing layer by layer to the output layer. The number of neurons in the output layer is the same as that in the input layer of the encoder. The decoder also employs the ReLU activation function in each hidden layer, while the output layer employs the Sigmoid activation function. The output value is mapped to the [0,1] interval through f(x) = 1 / (1+e^(-x)), matching the numerical range of the input features.

[0047] Next, the training samples were processed. An initial high-dimensional feature pool corresponding to different operating conditions and decay stages from historical data was selected as the training set. The training and validation sets were divided in a 7:3 ratio to ensure that the samples covered the characteristic changes throughout the entire life cycle of activated carbon. The training and validation set data were standardized so that the mean of each feature was 0 and the variance was 1, eliminating the dimensional differences between different feature dimensions and improving training stability. The processed training set data was then input into the autoencoder in batches, with each batch containing 32 samples to balance training efficiency and gradient stability.

[0048] Unsupervised training was then conducted. First, the network parameters were initialized. The weights of each layer were initialized using the Xavier initialization method, which randomly generates initial weights that conform to a uniform distribution based on the number of input and output neurons. The initial values ​​of the bias terms were all set to 0. The training process aimed to minimize the reconstruction error between the input features and the decoder output features, while incorporating sparsity constraints. The reconstruction error was calculated using the mean squared error formula. This is used to measure the difference between the reconstructed data and the original data. The sparsity constraint is achieved through an L1 regularization term, which adds a weighted sum of the absolute values ​​of the bottleneck layer activations to the loss function. The weight coefficient is set to 0.005, and the formula is: Total Loss = Reconstruction Error + 0.005 × Σ|Bottleneck Layer Activation Value|. This constraint forces the activation values ​​of most neurons in the bottleneck layer to approach 0, retaining only the activation of key neurons that are strongly correlated with the decline in adsorption performance.

[0049] During training, the Adam optimizer was used to adjust the network parameters, with a learning rate of 0.001 and momentum parameters β1=0.9 and β2=0.999. In each iteration, the feature compression process of the encoder was first calculated through forward propagation. The initial high-dimensional features were passed from the input layer to each hidden layer, and a nonlinear transformation was performed using the ReLU activation function, finally outputting a latent representation vector at the bottleneck layer. This latent representation vector was then passed to the decoder, undergoing inverse nonlinear transformation in each hidden layer and activation by the Sigmoid function in the output layer to obtain the reconstructed feature vector. The total loss was then calculated, and the weights and biases of each layer were updated backward along the network hierarchy using the backpropagation algorithm to reduce the total loss. The training iterations were set to 1000 epochs. After each epoch, the model performance was validated using a validation set. Training was stopped when the total loss on the validation set did not decrease for 20 consecutive epochs or decreased by less than 0.0001, at which point the network parameters reached their optimal state.

[0050] After training, all parameters of the encoder are fixed, and the initial high-dimensional feature pool data to be processed is input into the encoder. After nonlinear transformation and feature compression of the input layer and hidden layer, the latent representation vector output by the bottleneck layer is the core feature vector. This vector has automatically suppressed redundant features and only retained the key feature information that is strongly correlated with the degradation of activated carbon adsorption performance.

[0051] By clarifying the hierarchical structure, parameter settings, sample processing methods, and training process of the autoencoder, accurate dimensionality reduction and core feature selection of the initial high-dimensional feature pool were achieved, providing a high-quality feature foundation for the subsequent mapping of core feature vectors to the preset performance degradation scale.

[0052] Furthermore, the method provided in this application embodiment includes: The core feature vector is input into the preset performance degradation scale, and a scalar value between 0 and 1 is output as the phonon metabolic fingerprint; where 0 represents the initial new state and 1 represents the complete failure state that has reached the preset failure threshold.

[0053] Specifically, the first step is to construct a pre-defined performance degradation scale. Historical data of activated carbon throughout its entire lifecycle, from its initial virgin state to complete failure, is collected, with a total sample size of no less than 50 groups. This data should evenly cover five stages: virgin state, slight degradation, moderate degradation, severe degradation, and complete failure, with at least 8 samples for each stage. The data must include core feature vectors corresponding to different degradation stages, as well as the corresponding actual adsorption performance test results. The actual adsorption performance test indicators are set as dynamic breakthrough time and pollutant removal rate. Dynamic breakthrough time refers to the continuous operating time during which the pollutant concentration reaches the set outlet standard. The pollutant removal rate is calculated as (inlet pollutant concentration - outlet pollutant concentration) / inlet pollutant concentration × 100%.

[0054] The preset failure threshold is defined as the pollutant removal rate dropping to 80% of the initial virgin state removal rate, at which point the activated carbon is considered completely ineffective. Actual adsorption performance test results are divided into 10 levels according to the degree of failure, corresponding one-to-one with a scalar value range of 0 to 1. Completely satisfactory adsorption performance is defined as a removal rate ≥ 95% of the initial value and a dynamic breakthrough time ≥ 95% of the initial value, corresponding to a scalar value of 0. Adsorption performance dropping to the preset failure threshold corresponds to a scalar value of 1. The eight intermediate decay stages are evenly distributed across the scalar value range according to the proportion of adsorption performance decay, with each range spanning 0.1.

[0055] Next, a linear regression method is used to establish the mapping relationship between the core feature vector and the scalar value. The core feature vector from historical data is used as the input variable (features of each dimension are arranged in a fixed order), and the corresponding scalar value is used as the output variable. The regression coefficient is then solved using the least squares method, i.e., the coefficient that makes all samples... The coefficient matrix with the smallest sum. After construction, the significance of the mapping model needs to be tested, and the determination coefficients are required. The performance degradation scale must be at least 0.9; otherwise, historical samples are added and the model is retrained. The resulting mapping model with the preset performance degradation scale is stored as a function that can be directly called, ensuring the efficiency of the mapping calculation. At the same time, the 3σ principle is used to remove outliers from the historical data. That is, the mean and standard deviation of each feature are calculated. If any feature value of a sample exceeds the range of mean ± 3 × standard deviation, it is judged as an outlier and removed.

[0056] Next, consistency processing is performed on the input core feature vectors. First, the dimensions and data types of the core feature vectors are checked against the historical core feature vectors used when constructing the preset performance degradation scale. The vector dimensions must be identical, and the data type must be 64-bit floating-point. If dimension mismatches exist, features are filtered or supplemented according to the feature names and order used when constructing the scale. Features with the same historical feature names are retained, and missing features are supplemented using the median of the corresponding features from the same batch of historical data. If data types are inconsistent, they are uniformly converted to 64-bit floating-point using a numerical format conversion tool to ensure the compatibility between the input vector and the scale and avoid data conflicts during the mapping process. Simultaneously, outliers in the core feature vectors are identified and processed using the 3σ principle. The mean and standard deviation of each dimension of the input vector are calculated. If a feature value exceeds the range of mean ± 3 × standard deviation, it is considered an outlier and replaced with the median of that dimension's features to ensure the reliability of the input data.

[0057] Next, a mapping calculation is performed to synthesize the phonon metabolic fingerprint. The core feature vector after consistency processing is substituted into the mapping model with a preset performance degradation scale. The i-th eigenvalue of the core feature vector corresponds one-to-one with the i-th regression coefficient in the mapping model. Through the model's built-in linear regression calculation logic, the values ​​of each dimension of the core feature vector are multiplied with the corresponding regression coefficients and then summed to obtain the preliminary mapping result. The numerical precision is retained to 6 decimal places during the calculation process. After the calculation is completed, the result is subjected to boundary constraint processing. If the calculation result is less than 0, it is taken as 0; if it is greater than 1, it is taken as 1. If the result is in the range [0,1] but has a small fluctuation within ±0.001, no additional correction is required, and the original calculation result is directly retained to ensure that the output result is strictly within the scalar value range of 0 to 1.

[0058] Finally, the output scalar value is the phonon metabolic fingerprint. The closer the scalar value is to 0, the closer the microstructure of the activated carbon is to a brand-new state, and the better the adsorption performance. The closer the scalar value is to 1, the more severe the microstructure degradation of the activated carbon, and the closer the adsorption performance is to the failure threshold. Specifically, the scalar value in the range [0, 0.2) represents mild degradation, [0.2, 0.4) represents moderate degradation, [0.4, 0.6) represents moderate to severe degradation, [0.6, 0.8) represents severe degradation, and [0.8, 1.0] represents near failure or failure. This scalar value intuitively quantifies the current degradation state of the activated carbon, providing a clear quantitative basis for subsequent decay analysis and remaining service life prediction.

[0059] By constructing a mapping model through linear regression, verifying the consistency of input features, and performing linear weighted calculations and boundary constraints, the system achieves an accurate mapping of core feature vectors to scalar values ​​of 0 to 1, generating a phonon metabolic fingerprint that can intuitively reflect the degradation state of activated carbon, thus ensuring the reliability and practicality of the mapping results.

[0060] Furthermore, the method provided in this application embodiment includes: Based on the phonon spectrum data, the three-dimensional spatial distribution of acoustic parameters inside the activated carbon bed is inverted; the deterioration rate of acoustic parameters in local areas of the three-dimensional spatial distribution is calculated in real time; the second derivative of the phonon metabolic fingerprint change is calculated; when the deterioration rate or the second derivative exceeds a preset safety threshold, it is determined that there is a risk of sudden failure, and the highest level warning signal is generated to trigger emergency maintenance.

[0061] In one embodiment, the three-dimensional spatial distribution of acoustic parameters inside the activated carbon bed is retrieved based on phonon spectrum data. Specifically, according to the actual dimensions of the activated carbon bed (e.g., 1m × 1m × 1m), multiple microelectromechanical system (MEMS) sensor nodes are evenly embedded in the bed at 10cm intervals. A three-dimensional Cartesian coordinate system is established with the lower left corner of the bed as the origin, and the node coordinates are recorded. The sound velocity and attenuation coefficient are extracted from the preprocessed phonon spectrum data. The sound velocity is calculated as the ratio of the node spacing to the signal rise time, and the attenuation coefficient is obtained as the ratio of the peak value of the incident signal to the peak value of the received signal.

[0062] Next, a Gaussian radial basis function (RBF) was used for 3D spatial distribution inversion. The RBF shape parameter was set to 0.3. The bed was divided into grids with a side length of 5 cm, which is half the distance between sensor nodes. Using the coordinates of the sensor nodes and their corresponding acoustic parameters as known samples, the Euclidean distance between each grid node and all sensor nodes was calculated. These distances were then substituted into the Gaussian RBF to determine the weighting coefficients for each sample. The acoustic parameter values ​​for each grid node were obtained through weighted summation. All grid node data were integrated to form a complete 3D spatial distribution matrix of acoustic parameters. After the inversion, acoustic parameter sampling and detection were performed at the bed entity locations corresponding to five random grid nodes. The relative error between the detected values ​​and the inverted values ​​was required to be ≤5%. If the error exceeded this range, the RBF shape parameter was readjusted, and the inversion was performed again.

[0063] Then, the degradation rate of acoustic parameters in local areas is calculated in real time. A cubic local area with a side length of 10cm, consistent with the spacing between sensor nodes, is defined centered on each grid node. Overlap between these areas is allowed, but the overlap ratio is ≤50%. Time-series data is collected at intervals of once per minute. The weighted average of sound velocity and attenuation coefficient in each local area is calculated with a 50% weighting. The average is obtained by summing the sound velocity values ​​of all grid nodes in the area and dividing by the number of nodes, and summing the attenuation coefficient values ​​and dividing by the number of nodes, respectively. The initial value is calculated according to the formula: Degradation Rate = (Current Average - Previous Average) / Time Interval. After removing outliers using the 3σ principle, the average of three consecutive results is taken as the final degradation rate. A positive rate indicates acoustic parameter degradation; the larger the value, the more severe the degradation. The units for sound velocity are m / s·min, and the units for attenuation coefficient are dB / m·min.

[0064] Next, the second derivative of the phonon metabolic fingerprint change was calculated, and fingerprint scalar values ​​at least ten acquisition times were recorded simultaneously. Each acquisition time was strictly consistent with the acquisition time of the three-dimensional spatial distribution data of the acoustic parameters, i.e., the time difference between the acquisition time and the acoustic parameter acquisition time was ≤1 second. The first derivative was calculated using the central difference method at intermediate times, and the forward and backward difference methods were used to supplement the first and last times respectively. Then, the second derivative was calculated according to the same difference rule. Positive values ​​indicate accelerated deterioration, and negative values ​​indicate slowed deterioration.

[0065] The preset safety threshold is calibrated through a standardized process: collect more than 50 sets of activated carbon full life cycle data from normal operation to sudden failure, select 30 valid sudden failure cases, and statistically analyze the average rate of maximum acoustic parameter deterioration within 10 minutes before the failure in the cases. Take 80% of this average as the preset safety threshold for deterioration rate; statistically analyze the 95th percentile of the second derivative in the normal decay stage, and take 1.3 times this value as the preset safety threshold for the second derivative.

[0066] Finally, the final deterioration rate of each local area is compared in real time with the preset deterioration rate threshold under the corresponding operating condition, and the second derivative of the phonon metabolic fingerprint is compared with the preset second derivative threshold. If the deterioration rate of the sound velocity or attenuation coefficient of any local area exceeds the corresponding threshold, or the second derivative exceeds the preset threshold, and the warning signal is not lifted for 3 seconds, a sudden failure risk is immediately determined, and the highest-level warning signal containing the three-dimensional coordinates of the risk area, the current deterioration rate, the second derivative value, and operating condition information is generated. The warning signal triggers emergency maintenance: the control system automatically suspends the feeding operation, simultaneously starts the backup adsorption system to ensure the continuity of pollutant treatment, the on-site audible and visual alarms continue to sound, and at the same time, the warning information is pushed through the operation and maintenance management platform, and SMS notifications are sent to designated operation and maintenance personnel, requiring them to arrive on-site within 30 minutes and complete the emergency handling according to the process of risk location → sampling and testing → replacing activated carbon in the faulty area / repairing equipment.

[0067] By clearly defining core parameters and operating procedures, the system enables precise monitoring, rapid early warning, and efficient handling of sudden failure risks in activated carbon beds, ensuring that those skilled in the art can directly reproduce the problem and effectively guaranteeing the safety, stability, and continuity of the adsorption facility's operation.

[0068] Furthermore, the method provided in this application embodiment includes: The adsorption impact per unit time corresponding to the real-time operating condition data is analyzed to identify the operating condition mode. The real-time operating condition data includes the concentration of inlet pollutants, temperature, humidity, and pressure difference of the activated carbon bed. According to the operating condition mode, the capacity decay template characterizing the repeated adsorption-desorption of activated carbon is dynamically switched. The capacity decay template is used to predict the phonon metabolic fingerprint, and the predicted value of the remaining effective capacity corresponding to the operating condition mode and the predicted value of the remaining service life when the preset failure threshold is reached are output.

[0069] Optionally, real-time operating data of the activated carbon adsorption facility is first collected and processed. This real-time data includes inlet pollutant concentration, temperature, humidity, and activated carbon bed pressure differential. Data is collected at fixed time intervals of 1-5 minutes, with the specific interval adjusted according to the activated carbon decay rate: 1-2 minutes for faster decay and 3-5 minutes for slower decay, ensuring the data reflects changes in operating conditions promptly. The weighted summation method is used to calculate the adsorption impact value per unit time. The weights of each operating parameter are determined using the Pearson correlation coefficient method or grey relational analysis: with the activated carbon adsorption capacity decay rate as the dependent variable and each operating parameter as the independent variable, statistical analysis quantifies the correlation strength between each parameter and the decay rate. Stronger correlations result in higher weights, ensuring the weight allocation conforms to actual adsorption patterns.

[0070] Before calculation, a linear normalization method is used to map all operating parameters to the [0,1] interval. The specific formula is: Normalized value = (Measured parameter value - Minimum parameter value) / (Maximum parameter value - Minimum parameter value). The maximum and minimum parameter values ​​are taken from a reasonable range of historical operating data. Then, the values ​​are multiplied by their corresponding weights and summed to obtain the adsorption impact value per unit time. Based on the statistical analysis results of historical operating data, a critical threshold is set according to the quartiles or ternary quartiles of the adsorption impact value to divide the operating mode into three categories: light, medium, and heavy. For example, the impact value interval is divided into three segments according to the quartiles, corresponding to the three types of operating conditions. The operating mode identification is completed based on the interval to which the real-time calculated adsorption impact value belongs.

[0071] Next, activated carbon capacity matching is performed based on phonon metabolic fingerprints to accurately identify the current capacity decay starting point. Combining the identified operating mode with the determined capacity decay starting point, the capacity decay memory library is called to match the corresponding analog capacity decay curve set. By performing curve aggregation processing on this curve set, a capacity decay template characterizing the repeated adsorption-desorption of activated carbon is finally obtained. This step will be explained in detail in the following content.

[0072] Finally, the actual adsorption capacity corresponding to the current phonon metabolic fingerprint is used as the starting capacity and substituted into the capacity decay template. The process is then extrapolated along the template curve trajectory using time steps consistent with the curve's time axis, such as 1 hour / step or 24 hours / step. The preset failure threshold capacity is determined based on pollutant emission compliance requirements and critical adsorption efficiency values: first, the pollutant emission concentration standards that the adsorption facility must meet are clearly defined; then, the minimum adsorption efficiency corresponding to compliance is experimentally calibrated; finally, this adsorption efficiency is converted into the corresponding activated carbon adsorption capacity, which is the preset failure threshold capacity. Based on the capacity change over time pattern in the template curve, the capacity change interval between the current starting capacity and the preset failure threshold capacity is read. The capacity difference within this interval is the predicted value of the remaining effective capacity. Simultaneously, the time from the current point in time to the point where the capacity drops to the preset failure threshold capacity is statistically analyzed. This time length is the predicted value of the remaining service life before reaching the preset failure threshold. The prediction results are output according to maintenance decision requirements: daily maintenance is accurate to the hour, and planned replacement is accurate to the day.

[0073] Furthermore, a deviation correction mechanism needs to be added during the extrapolation process: if the actual adsorption capacity corresponding to the real-time collected phonon metabolic fingerprint deviates from the concurrent capacity predicted by the template by more than ±10%, the analog curve set is re-screened and the capacity decay template is updated to ensure prediction accuracy; if there is missing operating condition data, the missing data is filled by the historical concurrent average or interpolation method; if there is no matching analog curve in the capacity decay memory, the curve set with similar operating conditions and similar decay start points in the library is called, and after fine-tuning in combination with the current operating condition characteristics, it is aggregated, or the default general decay template is enabled, which can be constructed based on the average decay law of the entire operating condition.

[0074] Through the above-mentioned sequential steps, combined with phonon metabolic fingerprints and real-time operating data, the remaining effective capacity and remaining service life of activated carbon are accurately output, providing a clear and feasible quantitative basis for the preventive maintenance of adsorption facilities, and ensuring the scientific and timely nature of maintenance decisions.

[0075] Furthermore, the method provided in this application embodiment includes: Activated carbon capacity matching is performed based on the phonon metabolic fingerprint to identify the current capacity decay starting point; based on the operating mode and the capacity decay starting point, the capacity decay memory is called to match the analog capacity decay curve set, and the analog capacity decay curve set is aggregated to obtain the capacity decay template.

[0076] In one embodiment, firstly, considering that the decay rate may differ at different decay nodes even under the same operating conditions, to avoid subsequent capacity decay template matching deviations due to misjudgment of decay nodes, it is necessary to first establish the correspondence between the phonon metabolic fingerprint (0-1 scalar value) and the actual adsorption capacity of activated carbon using no fewer than 30 sets of historical data covering different decay stages. A conversion model is then constructed using linear or nonlinear fitting (such as polynomial fitting), and the fitting equation must satisfy the coefficient of determination. To ensure fitting accuracy, the model directly converts phonon metabolic fingerprint scalar values ​​into corresponding actual adsorption capacities. The current phonon metabolic fingerprint scalar value is then substituted into the conversion model to calculate the current actual adsorption capacity, which is compared with the measured initial adsorption capacity of virgin activated carbon. The measured initial adsorption capacity is calibrated through static adsorption experiments or dynamic breakthrough experiments, with experimental conditions consistent with actual operating conditions. Subsequently, by tracking the capacity decay curve throughout the activated carbon's life cycle, the inflection point where adsorption capacity transitions from gradual decay to accelerated decay is identified. The proportion of the capacity corresponding to this inflection point relative to the initial adsorption capacity is used as the critical decay ratio, for example, 85%-90%. If the current actual adsorption capacity drops to this critical ratio, it is determined as the current capacity decay starting point; if the current actual adsorption capacity is already below this ratio, the state corresponding to the current capacity is directly used as the decay starting point, ensuring that the subsequently called capacity decay template accurately matches the current actual decay node.

[0077] Next, combining the identified operating conditions and the determined capacity decay starting point, the capacity decay memory is used to match the corresponding set of analog capacity decay curves. The capacity decay memory needs to be pre-built, and each capacity decay curve stored in the memory must meet the following criteria: "complete decay cycle + clear operating condition label + clear decay starting point label." The curve data comes from the measured capacity time-series data of activated carbon from the decay starting point to failure under the corresponding operating condition and decay starting point. The time axis unit is consistent with the operating condition data acquisition interval, and the capacity unit is consistent with the actual adsorption capacity. During matching, dynamic time warping (DTW) or Euclidean distance is used to set a similarity threshold. For example, a similarity ≥ 85% is considered a match. All analog capacity decay curves that match the current operating condition and decay starting point are selected to form a set of analog capacity decay curves. The curve set is then subjected to curve aggregation processing: Capacity data points for each curve are extracted at a fixed time step consistent with the data acquisition interval of the operating conditions. After removing outlier data points that exceed the mean ±3σ, the arithmetic mean or weighted mean of the capacity data of all valid curves at each time point is calculated. The weight is the similarity between the curve and the current operating condition. A smooth aggregated curve is plotted with time as the x-axis and mean capacity as the y-axis. This aggregated curve is the capacity attenuation template corresponding to the current operating condition mode and the attenuation starting point.

[0078] In summary, the activated carbon adsorption facility maintenance method based on lifespan prediction provided in this application has the following technical effects: This application utilizes an acoustic sensor array within an activated carbon bed to emit high-frequency sound waves and receive response signals to obtain phonon spectrum data. Multi-dimensional features are extracted and dimensionality-reduced to construct a phonon metabolic fingerprint. Combined with real-time operating condition analysis of decay and matching attenuation templates, the remaining lifetime is predicted. Simultaneously, it monitors the risk of sudden failure, enabling preventative maintenance of activated carbon adsorption facilities and improving the accuracy and reliability of maintenance. This achieves the technical effect of accurately predicting the lifetime of activated carbon adsorption facilities and providing early warning of sudden risks.

[0079] Example 2, as Figure 2 As shown, based on the same inventive concept as in Embodiment 1 above, this application provides an activated carbon adsorption facility maintenance system based on lifespan prediction, the system comprising: Phonon spectrum data acquisition module 1 is used to transmit high-frequency sound wave signals to the activated carbon adsorption facility through an acoustic sensor array deployed in the activated carbon adsorption facility, and receive response signals to obtain phonon spectrum data reflecting the lattice vibration state of activated carbon.

[0080] Phonon metabolic fingerprint construction module 2 is used to extract characteristic parameters representing changes in the microstructure state of activated carbon from the phonon spectrum data and construct a dynamically evolving phonon metabolic fingerprint. The phonon metabolic fingerprint is a low-dimensional feature vector obtained by dimensionality reduction after multi-dimensional analysis of the phonon spectrum data.

[0081] The remaining service life prediction value acquisition module 3 is used to combine the phonon metabolic fingerprint and the real-time operating condition data of the activated carbon adsorption facility to perform activated carbon decay analysis and output the remaining service life prediction value of the activated carbon when it reaches the preset failure threshold.

[0082] Preventive maintenance execution module 4 performs preventive maintenance based on the remaining useful life prediction value.

[0083] Furthermore, the phonon metabolic fingerprint construction module 2 is used to perform the following steps: Frequency domain features, time domain features, and spatial correlation features are extracted from the phonon spectrum data to form an initial high-dimensional feature pool; an autoencoder is used to reduce the dimensionality and select features from the initial high-dimensional feature pool to obtain a core feature vector; the core feature vector is mapped to a preset performance attenuation scale to synthesize the phonon metabolic fingerprint.

[0084] Furthermore, the phonon metabolic fingerprint construction module 2 is used to perform the following steps: Frequency domain features include phonon density of states peak parameters, spectral moment characteristics, and resonant trajectory; time domain features include phonon lifetime distribution and waveform distortion; spatial correlation features include heterogeneity index and correlation length.

[0085] Furthermore, the phonon metabolic fingerprint construction module 2 is used to perform the following steps: An autoencoder neural network comprising an encoder and a decoder is constructed. An initial high-dimensional feature pool from historical data is used as training samples. The goal is to minimize the reconstruction error between the input features and the decoder output features. The autoencoder neural network is trained unsupervised. The encoder compresses the initial high-dimensional feature pool into a bottleneck layer through multiple nonlinear transformations. The output of neurons in the bottleneck layer is a latent representation vector that integrates the original multi-feature information. During training, sparsity constraints on the activation values ​​of the bottleneck layer are added to the loss function, causing the autoencoder neural network to activate key neurons strongly correlated with the decay of adsorption performance. This automatically suppresses redundant features in the latent representation vector, completing the reconstruction of the core feature vector.

[0086] Furthermore, the phonon metabolic fingerprint construction module 2 is used to perform the following steps: The core feature vector is input into the preset performance degradation scale, and a scalar value between 0 and 1 is output as the phonon metabolic fingerprint; where 0 represents the initial new state and 1 represents the complete failure state that has reached the preset failure threshold.

[0087] Furthermore, the remaining useful life prediction value acquisition module 3 is used to perform the following steps: The adsorption impact per unit time corresponding to the real-time operating condition data is analyzed to identify the operating condition mode. The real-time operating condition data includes the concentration of inlet pollutants, temperature, humidity, and pressure difference of the activated carbon bed. According to the operating condition mode, the capacity decay template characterizing the repeated adsorption-desorption of activated carbon is dynamically switched. The capacity decay template is used to predict the phonon metabolic fingerprint, and the predicted value of the remaining effective capacity corresponding to the operating condition mode and the predicted value of the remaining service life when the preset failure threshold is reached are output.

[0088] Furthermore, the remaining useful life prediction value acquisition module 3 is used to perform the following steps: Activated carbon capacity matching is performed based on the phonon metabolic fingerprint to identify the current capacity decay starting point; based on the operating mode and the capacity decay starting point, the capacity decay memory is called to match the analog capacity decay curve set, and the analog capacity decay curve set is aggregated to obtain the capacity decay template.

[0089] Furthermore, the phonon metabolic fingerprint construction module 2 is used to perform the following steps: Based on the phonon spectrum data, the three-dimensional spatial distribution of acoustic parameters inside the activated carbon bed is inverted; the deterioration rate of acoustic parameters in local areas of the three-dimensional spatial distribution is calculated in real time; the second derivative of the phonon metabolic fingerprint change is calculated; when the deterioration rate or the second derivative exceeds a preset safety threshold, it is determined that there is a risk of sudden failure, and the highest level warning signal is generated to trigger emergency maintenance.

[0090] Furthermore, the phonon spectrum data acquisition module 1 is used to perform the following steps: The acoustic sensor array is composed of multiple microelectromechanical system (MEMS) sensor nodes uniformly embedded inside the activated carbon bed of the activated carbon adsorption facility; the multiple MEMS sensor nodes transmit sound waves based on a preset frequency and receive response signals to obtain the phonon spectrum data.

[0091] The activated carbon adsorption facility maintenance system based on lifespan prediction provided in this embodiment of the invention can execute the activated carbon adsorption facility maintenance method based on lifespan prediction provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0092] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0093] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A maintenance method for activated carbon adsorption facilities based on lifespan prediction, characterized in that, include: By using an acoustic sensor array deployed in the activated carbon adsorption facility, high-frequency acoustic signals are emitted into the activated carbon adsorption facility, and response signals are received to obtain phonon spectrum data reflecting the lattice vibration state of activated carbon. Characteristic parameters representing changes in the microstructure of activated carbon are extracted from the phonon spectrum data to construct a dynamically evolving phonon metabolic fingerprint. The phonon metabolic fingerprint is a low-dimensional feature vector obtained by dimensionality reduction after multi-dimensional analysis of the phonon spectrum data. By combining the phonon metabolic fingerprint and the real-time operating data of the activated carbon adsorption facility, the decay analysis of the activated carbon is performed, and the predicted remaining service life of the activated carbon when it reaches the preset failure threshold is output. Preventative maintenance is performed based on the predicted remaining useful life.

2. The method for maintaining activated carbon adsorption facilities based on lifespan prediction as described in claim 1, characterized in that, Characteristic parameters representing changes in the microstructure of activated carbon are extracted from the phonon spectrum data to construct a dynamically evolving phonon metabolic fingerprint, including: Frequency domain features, time domain features, and spatial correlation features are extracted from the phonon spectrum data to form an initial high-dimensional feature pool. The initial high-dimensional feature pool is reduced in dimensionality and feature selection is performed using an autoencoder to obtain the core feature vector; The core feature vector is mapped to a preset performance degradation scale to synthesize the phonon metabolic fingerprint.

3. The method for maintaining activated carbon adsorption facilities based on lifespan prediction as described in claim 2, characterized in that, Frequency domain features include phonon density of states peak parameters, spectral moment characteristics, and resonance trajectory; Temporal characteristics include phonon lifetime distribution and waveform distortion; spatial correlation characteristics include heterogeneity index and correlation length.

4. The method for maintaining activated carbon adsorption facilities based on lifespan prediction as described in claim 2, characterized in that, The initial high-dimensional feature pool is reduced in dimensionality and feature selection is performed using an autoencoder to obtain the core feature vector, including: An autoencoder neural network containing encoder and decoder parts is constructed. An initial high-dimensional feature pool from historical data is used as training samples. The autoencoder neural network is trained unsupervised with the goal of minimizing the reconstruction error between the input features and the decoder output features. The encoder compresses the initial high-dimensional feature pool into a bottleneck layer through multi-layer nonlinear transformation. The output of the neurons in the bottleneck layer is a potential representation vector that integrates the original multi-feature information. During training, by adding sparsification constraints on the activation values ​​of the bottleneck layer to the loss function, the autoencoder neural network activates key neurons that are strongly correlated with the decay of adsorption performance. In the potential representation vector, redundant features are automatically suppressed, and the core feature vector is reconstructed.

5. The method for maintaining activated carbon adsorption facilities based on lifespan prediction as described in claim 2, characterized in that, The core feature vector is mapped to a preset performance degradation scale to synthesize the phonon metabolic fingerprint, including: The core feature vector is input into the preset performance degradation scale, and a scalar value between 0 and 1 is output as the phonon metabolic fingerprint; where 0 represents the initial new state and 1 represents the complete failure state that has reached the preset failure threshold.

6. The method for maintaining activated carbon adsorption facilities based on lifespan prediction as described in claim 1, characterized in that, Combining the phonon metabolic fingerprint with the real-time operating data of the activated carbon adsorption facility, a decay analysis of the activated carbon is performed, outputting a predicted remaining service life of the activated carbon when it reaches a preset failure threshold, including: Analyze the adsorption impact per unit time corresponding to the real-time operating condition data to identify the operating mode. The real-time operating condition data includes the concentration of inlet pollutants, temperature, humidity, and pressure difference of the activated carbon bed. Based on the operating mode, dynamically switch the capacity decay template characterizing the repeated adsorption-desorption of activated carbon. The phonon metabolism fingerprint is predicted using the capacity decay template, and the predicted value of the remaining effective capacity corresponding to the operating mode and the predicted value of the remaining service life when the preset failure threshold is reached are output.

7. The method for maintaining activated carbon adsorption facilities based on lifespan prediction as described in claim 6, characterized in that, Based on the aforementioned operating mode, the capacity decay template characterizing repeated adsorption-desorption of activated carbon is dynamically switched, including: Activated carbon capacity matching is performed based on the phonon metabolic fingerprint to identify the current capacity decay starting point; Based on the operating mode and the capacity decay starting point, the capacity decay memory is called to match the analog capacity decay curve set, and the analog capacity decay curve set is aggregated to obtain the capacity decay template.

8. The method for maintaining activated carbon adsorption facilities based on lifespan prediction as described in claim 1, characterized in that, After constructing the dynamically evolving phonon metabolic fingerprint, the following steps are also included: Based on the phonon spectrum data, the three-dimensional spatial distribution of acoustic parameters inside the activated carbon bed is retrieved. Real-time calculation of the rate of deterioration of acoustic parameters in local areas within the three-dimensional spatial distribution; Calculate the second derivative of the changes in the phonon metabolic fingerprint; When the deterioration rate or the second derivative exceeds a preset safety threshold, a sudden failure risk is determined, and the highest-level warning signal is generated to trigger emergency maintenance.

9. The method for maintaining activated carbon adsorption facilities based on lifespan prediction as described in claim 1, characterized in that, By using an acoustic sensor array deployed in the activated carbon adsorption facility, high-frequency acoustic signals are emitted into the activated carbon adsorption facility, and response signals are received to obtain phonon spectrum data reflecting the lattice vibration state of the activated carbon, including: The acoustic sensor array is composed of multiple microelectromechanical system (MEMS) sensor nodes that are evenly distributed and embedded inside the activated carbon bed of the activated carbon adsorption facility. The multiple microelectromechanical system sensor nodes transmit sound waves at a preset frequency and receive response signals to obtain the phonon spectrum data.

10. A maintenance system for activated carbon adsorption facilities based on lifespan prediction, characterized in that, For implementing the activated carbon adsorption facility maintenance method based on lifespan prediction as described in any one of claims 1-9, the system comprises: The phonon spectrum data acquisition module is used to transmit high-frequency sound wave signals to the activated carbon adsorption facility through an acoustic sensor array deployed in the activated carbon adsorption facility, and receive response signals to obtain phonon spectrum data reflecting the lattice vibration state of activated carbon. The phonon metabolic fingerprint construction module is used to extract characteristic parameters representing changes in the microstructure state of activated carbon from the phonon spectrum data and construct a dynamically evolving phonon metabolic fingerprint. The phonon metabolic fingerprint is a low-dimensional feature vector obtained by dimensionality reduction after multi-dimensional analysis of the phonon spectrum data. The remaining service life prediction module is used to combine the phonon metabolic fingerprint and the real-time operating data of the activated carbon adsorption facility to perform activated carbon decay analysis and output the remaining service life prediction value of the activated carbon when it reaches the preset failure threshold. The preventive maintenance execution module performs preventive maintenance based on the predicted remaining useful life.

Images