Intelligent monitoring method and system for sand and gravel concrete refrigeration equipment

Abstract

This application discloses an intelligent monitoring method and system for sand and gravel concrete refrigeration equipment, relating to the field of intelligent equipment monitoring. It utilizes dual-channel sensors deployed in parallel inside and outside the air duct to collect mixed signals including airflow sound and ambient background noise. Through timestamp synchronization and adaptive background noise cancellation processing, the pure airflow friction sound signal is effectively extracted. Furthermore, specific modulation is generated on the high-frequency acoustic characteristics of the flowing air, and the acoustic roughness index is calculated by quantizing and extracting the microscopic roughness acoustic fingerprint from the high-frequency acoustic fingerprint feature spectrum. Based on this, the system monitors the acoustic fingerprint evolution trajectory, accurately determines the current frosting phase state, and generates optimized calibration control commands at the initial stage of frosting, realizing proactive defensive adjustment of the refrigeration unit, thereby solving the problems of monitoring failure and defrosting control lag in existing technologies.

Description

Technical Field

[0001] This application relates to the field of intelligent equipment monitoring, and more specifically, to an intelligent monitoring method and system for sand and gravel concrete refrigeration equipment. Background Technology

[0002] Sand and gravel concrete refrigeration equipment plays a crucial role in the production of high-performance concrete. Its core function is to control the outlet temperature of concrete by reducing the temperature of aggregates, thereby effectively preventing temperature cracking in large-volume concrete. Due to the unique characteristics of sand and gravel cooling operations, the equipment operates in a harsh environment with high dust (such as cement and stone powder) and high humidity for extended periods. As the core component of heat exchange, the evaporator is highly susceptible to surface frost and even ice formation. Once frost accumulates, it not only creates thermal resistance, significantly reducing heat exchange efficiency, but can also block air ducts in severe cases, leading to unit shutdown and directly impacting the continuity and quality stability of concrete production. Therefore, developing an intelligent monitoring system for the operating status of this type of equipment, especially for the evaporator frost formation process, is essential for ensuring system energy efficiency and production safety.

[0003] However, existing monitoring technologies mainly rely on macroscopic threshold triggering mechanisms, such as differential pressure switch alarms or preset fixed-time defrosting strategies, which generally suffer from significant lag and limitations. On the one hand, traditional contact or invasive sensors (such as ice thickness probes and optical icing sensors) are difficult to apply in the unique environment of sand and gravel concrete plants. They are easily covered by mud and dust mixed with moisture and become ineffective. Furthermore, they are difficult to install inside the compact finned tube evaporator, making it impossible to directly observe the microscopic icing state. On the other hand, monitoring methods based on airflow differential pressure are essentially passive identification methods. An alarm is only triggered when the ice layer accumulates enough to severely obstruct airflow and cause a sharp increase in fan differential pressure. At this point, heat exchange efficiency has often been significantly reduced, and the system has actually been operating with defects for a long time. The energy consumption and time required for subsequent defrosting are extremely high. On the other hand, timed defrosting methods do not consider actual operating conditions such as air humidity, which can easily lead to frequent defrosting, wasting energy, or causing unnecessary fluctuations in storage temperature in low-humidity environments. Currently, there is a lack of non-invasive monitoring methods that can accurately capture the characteristics of the early stage of micro-icing (frost crystal nucleation stage) without contacting high-dust surfaces, making it difficult to achieve early and accurate identification of the frost status of refrigeration equipment and proactive defensive adjustment. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this application provides an intelligent monitoring method for sand and gravel concrete refrigeration equipment, comprising: S1: acquiring an internal mixed signal containing airflow sound and background transmitted sound, and an external reference signal containing only environmental background noise, through monitoring channel sensors deployed in parallel within the refrigeration unit's air duct and a reference channel sensor deployed outside the unit's chassis; S2: performing time-stamp synchronization alignment and pre-emphasis processing on the internal mixed signal and the external reference signal to obtain a dual-channel synchronous audio dataset; S3: performing adaptive background noise cancellation on the dual-channel synchronous audio dataset to obtain a pure audio dataset. S4: Perform short-time Fourier transform and frequency domain mapping on the clean airflow friction sound signal to obtain a high-frequency acoustic fingerprint feature spectrum; S5: Quantize and extract the micro-roughness acoustic fingerprint from the high-frequency acoustic fingerprint feature spectrum to obtain the acoustic roughness index; S6: Based on the preset system cleaning benchmark fingerprint, intelligently determine the frosting state based on acoustic fingerprint evolution to determine the current frosting phase state of the evaporator; S7: Based on the frosting phase state, generate an optimized calibration control command, which is sent to the refrigeration unit controller to perform active defensive adjustment.

[0005] This application also provides an intelligent monitoring system for a sand and gravel concrete refrigeration equipment, comprising: a signal acquisition module for acquiring an internal mixed signal containing airflow sound and background transmitted sound, and an external reference signal containing only environmental background noise, through monitoring channel sensors deployed in parallel within the refrigeration unit's air duct and a reference channel sensor deployed outside the unit's chassis; a signal synchronization module for performing time-stamp synchronization alignment and pre-emphasis processing on the internal mixed signal and the external reference signal to obtain a dual-channel synchronized audio dataset; a background noise cancellation module for adaptively canceling background noise on the dual-channel synchronized audio dataset to obtain a pure airflow friction sound signal; and high-frequency acoustic feature extraction. The system includes a sound extraction module for performing short-time Fourier transform and frequency domain mapping on the pure airflow friction sound signal to obtain a high-frequency acoustic fingerprint feature spectrum; an acoustic roughness analysis module for quantizing and extracting the microscopic roughness acoustic fingerprint from the high-frequency acoustic fingerprint feature spectrum to obtain an acoustic roughness index; a frosting state discrimination module for intelligently discriminating the frosting state based on acoustic fingerprint evolution of the acoustic roughness index to determine the current frosting phase state of the evaporator based on a preset system cleaning benchmark fingerprint; and a control command generation and distribution module for generating optimized calibration control commands based on the frosting phase state, which are then distributed to the refrigeration unit controller to perform proactive defensive adjustments.

[0006] Compared with existing technologies, this application provides an intelligent monitoring method and system for sand and gravel concrete refrigeration equipment, addressing the technical pain point of difficulty in early detection of evaporator frost formation in high-dust and high-humidity environments. This solution does not rely on easily contaminated contact probes or lag-prone differential pressure switches, but instead utilizes dual-channel sensors deployed in parallel inside and outside the air duct to collect mixed signals including airflow sound and ambient background noise. Through timestamp synchronization and adaptive background noise cancellation processing, the pure airflow friction sound signal is effectively extracted. Given that the nucleation and accumulation of frost crystals on the evaporator surface microscopically alters the physical roughness of the fins, thereby specifically modulating the high-frequency acoustic characteristics of the flowing air, the acoustic roughness index is calculated by quantifying and extracting the microscopic roughness acoustic fingerprint from the high-frequency acoustic fingerprint feature spectrum. Based on this, the system monitors the acoustic fingerprint evolution trajectory, accurately determines the current frost phase state, and generates optimized calibration control commands in the early stages of frost formation, achieving proactive defensive adjustment of the refrigeration unit, thus solving the problems of monitoring failure and delayed defrosting control in existing technologies. Attached Figure Description

[0007] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings.

[0008] Figure 1 This is a flowchart of an intelligent monitoring method for a sand and gravel concrete refrigeration equipment according to an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of the data flow of an intelligent monitoring method for a sand and gravel concrete refrigeration equipment according to an embodiment of this application.

[0010] Figure 3 This is a flowchart of step S5 in the intelligent monitoring method for sand and gravel concrete refrigeration equipment according to an embodiment of this application.

[0011] Figure 4 This is a block diagram of an intelligent monitoring system for a sand and gravel concrete refrigeration equipment according to an embodiment of this application. Detailed Implementation

[0012] The embodiments of this application will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0013] In view of the shortcomings in the above-mentioned technical field, this application proposes an intelligent monitoring method for sand and gravel concrete refrigeration equipment. Figure 1 This is a flowchart of an intelligent monitoring method for a sand and gravel concrete refrigeration equipment according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in an intelligent monitoring method for a sand and gravel concrete cooling equipment according to an embodiment of this application. Figure 1 and Figure 2 As shown, the intelligent monitoring method for a sand and gravel concrete refrigeration equipment according to an embodiment of this application includes: S1: acquiring an internal mixed signal containing airflow sound and background transmitted sound, and an external reference signal containing only environmental background noise, through monitoring channel sensors deployed in parallel within the refrigeration unit's air duct and a reference channel sensor deployed outside the unit's chassis; S2: performing time-stamp synchronization alignment and pre-emphasis processing on the internal mixed signal and the external reference signal to obtain a dual-channel synchronous audio dataset; S3: performing adaptive background noise cancellation on the dual-channel synchronous audio dataset to obtain pure airflow friction sound. S4: Perform short-time Fourier transform and frequency domain mapping processing on the pure airflow friction sound signal to obtain a high-frequency acoustic fingerprint feature spectrum; S5: Quantize and extract the micro-roughness acoustic fingerprint from the high-frequency acoustic fingerprint feature spectrum to obtain the acoustic roughness index; S6: Based on the preset system cleaning benchmark fingerprint, intelligently determine the frosting state based on acoustic fingerprint evolution to determine the current frosting phase state of the evaporator; S7: Based on the frosting phase state, generate an optimized calibration control command, which is sent to the refrigeration unit controller to perform active defensive adjustment.

[0014] In step 1, monitoring channel sensors deployed in parallel within the refrigeration unit's air duct and reference channel sensors deployed outside the unit's chassis are used to collect an internal mixed signal containing airflow noise and background transmitted sound, as well as an external reference signal containing only environmental background noise. It should be understood that in the actual operating conditions of sand and gravel concrete refrigeration equipment, the plant environment is extremely harsh, typically accompanied by strong background noise from crushers, mixing towers (the main vibration source), and heavy transport vehicles. This broadband, high-intensity environmental noise not only has a high sound pressure level but also possesses extremely strong penetrating power, capable of penetrating through the metal chassis walls of the unit and entering the air duct, where it overlaps with the weak airflow friction sound generated by the airflow passing through the evaporator fin gaps. Because the change in airflow acoustic signature caused by the change in fin surface roughness due to frost crystal nucleation in the early stages of frosting is extremely subtle, in single-channel acquisition mode, this weak characteristic signal is easily submerged by the high signal-to-noise ratio of industrial background noise, leading to the failure of subsequent feature extraction and state determination. In order to accurately extract the pure airflow acoustic features that characterize the frosting state, this application uses monitoring channel sensors deployed in parallel inside the air duct of the refrigeration unit and reference channel sensors deployed outside the unit chassis to simultaneously collect acoustic data inside and outside the air duct. This results in an internal mixed signal containing airflow sound and background transmitted sound, as well as an external reference signal containing only environmental background noise, providing a homogeneous data basis for subsequent algorithms to perform differential cancellation.

[0015] In one embodiment of step S1, the specific processing is as follows: The implementation of this step relies on a precisely arranged acoustic sensor array and a high-fidelity signal acquisition link. During implementation, the optimal deployment points of the sensors must first be determined within the physical architecture of the refrigeration unit. The monitoring channel sensor uses an industrial-grade electret microphone or MEMS microphone with high sensitivity and wide frequency response characteristics. It is rigidly mounted inside the air duct on the return air side of the evaporator, with the probe pointing away from the fan blade rotation area to reduce direct wind noise interference, focusing on capturing the aerodynamic noise as the airflow passes through the evaporator fin array. The reference channel sensor uses the same type of acoustic sensor as the monitoring channel sensor, deployed on the outer casing of the unit chassis. The location must avoid the unit's own cooling fan exhaust vents and directly face the core noise source of the sand and gravel plant area, such as the mixing main building or the aggregate conveyor belt side, to ensure maximum pickup of pure environmental background noise. During equipment operation, the sensors of both channels work in parallel, continuously converting the received sound pressure fluctuations into analog voltage signals.

[0016] At this point, the monitoring channel sensor collects the internal mixed signal within the enclosed air duct. This signal is essentially a composite analog voltage waveform, physically composed of two main components: the first is airflow noise, which is high-frequency aerodynamic noise generated by the shearing action of the fluid boundary layer and the friction between the fluid and the solid wall as cold air flows at a specific velocity across the evaporator fin surface. This component is directly controlled by the geometry of the evaporator surface, directly reflecting the frost accumulation state. The second component is background transmitted sound, which is the residual noise component of mechanical noise from the external environment that still penetrates into the air duct after being attenuated by the sound insulation of the chassis wall. Simultaneously, the external reference channel sensor synchronously collects the external reference signal. This signal is the raw analog voltage signal containing only background ambient noise. Because the external sensor is physically isolated from the airflow inside the air duct and exposed to an open environment, the recorded waveform only reflects the superposition of external sound sources such as crushers and mixers within the plant area, and does not contain any information about the friction of the airflow inside the evaporator.

[0017] The raw analog voltage signals acquired from these two channels are transmitted in parallel to the front-end interface of the signal processing unit via shielded twisted-pair cables. To prevent electromagnetic interference during signal transmission, the transmission lines require strict grounding and shielding measures. Upon arrival at the acquisition end, these two continuous analog signals are used as the raw input to the algorithm and are labeled as internal mixed signal channel data and external reference signal channel data, respectively. At this stage, complex digital filtering or frequency domain transformation is not involved; the focus is on high-fidelity signal acquisition. For example, in practical operation, if the external ambient noise is mainly concentrated in the low-frequency band below 500Hz, while the high-frequency characteristics of airflow friction sound extend to above 15kHz, then the frequency response range of the sensor and subsequent acquisition circuits must cover at least the full frequency band from 20Hz to 20kHz. The signal acquired at this moment has not yet been discretized and is represented by a voltage amplitude that changes continuously over time. For example, the internal mixed signal may be represented by a peak-to-peak fluctuation of several hundred millivolts superimposed on a 2.5V DC bias, which contains both low-frequency rumbling and high-frequency hissing. The external reference signal is mainly represented by an external noise waveform that is highly correlated with the low-frequency components of the internal signal.

[0018] In step S2, the internal mixed signal and the external reference signal are time-stamped and pre-emphasized to obtain a dual-channel synchronous audio dataset. Correspondingly, in the signal acquisition process of the previous stage, although the acquired internal mixed signal and external reference signal accurately record the physical fluctuations of their respective sound fields, they are essentially analog voltage waveforms. Furthermore, due to the limited speed of sound propagation in air and the difference in the physical spatial positions of the two sensors, there will inevitably be a timing deviation in the arrival time of the same background noise source at the two sensors. If the signal with a phase difference is directly used for subsequent differential noise reduction, it will not only fail to effectively cancel the background noise but may also cause a multiplication of noise energy due to phase misalignment. In addition, the mechanical vibration of the refrigeration unit during operation is mainly concentrated in the low-frequency band, with extremely high energy density, while the airflow friction sound characterizing early frost formation is distributed in the high-frequency band and has weak energy. This significant spectral tilt makes the high-frequency characteristics easily masked by low-frequency pink noise. Therefore, in order to construct a mathematically rigorous and feature-saliency balanced analytical benchmark, this application performs precise time-domain synchronization and frequency-domain energy reshaping on the original signal before performing core noise reduction processing. Through timestamp synchronization alignment and pre-emphasis processing, the analog waveform is transformed into a computer-processable, time-consistent, and high-frequency enhanced dual-channel synchronous audio dataset.

[0019] In one embodiment, step S2 includes: S21: performing parallel discretization and quantization on the internal mixed signal and the external reference signal to obtain an internal discrete sequence and an external discrete sequence; S22: performing time delay estimation and alignment on the internal discrete sequence and the external discrete sequence based on the cross-correlation function to obtain an aligned external sequence; S23: performing spectral tilt compensation on the internal discrete sequence and the aligned external sequence to obtain a dual-channel synchronous audio dataset.

[0020] The specific processing is as follows: First, step S21 is executed. The input of this step directly receives the two continuous analog voltage signals output by the sensor in step S1. During implementation, the signal flow first enters the front end of a high-precision multi-channel analog-to-digital converter (ADC). Considering that the airflow friction sound generated by early frost in sand and gravel concrete refrigeration equipment has wide bandwidth characteristics, its effective information often extends to above 10kHz. In order to fully preserve these high-frequency microscopic details and follow the Nyquist sampling theorem, the sampling rate setting needs to reserve sufficient redundancy. In the specific configuration, the sampling rate is set to 48kHz, which means that the analog signal is captured 48,000 times per second for instantaneous voltage values. The quantization circuit inside the ADC adopts a parallel triggering mechanism, that is, on the rising edge of the same clock pulse, the voltage of the internal channel and the external channel are held and quantized simultaneously. If the quantization depth is 24 bits, the voltage value of each sampling point is mapped to a value between 10kHz and 24kHz. arrive Integers between [a certain value]. This process is mathematically described by a discretization calculation formula: for an internally mixed signal, the discrete value of its nth sampling point is [a certain value]. Represented as Where Signal InternalRaw represents the original analog voltage function, The sampling period is 1 / 48000 of a second. This represents the quantization mapping function. Similarly, the external reference signal is also synchronously converted into a discrete sequence. In actual data stream processing, these continuously generated discrete values ​​are not processed one by one, but are encapsulated into a fixed-length time window. For example, if the frame length is set to 4096 sampling points, then each frame of data represents approximately 85 milliseconds of physical duration. After this processing step, the originally continuous analog waveform is transformed into two sets of digital vectors of the same length with one-to-one time indices, i.e., the internal discrete sequence. and external discrete sequence .

[0021] Following step S22, in the physical scenario, the external reference sensor is deployed on the chassis shell, directly exposed to environmental noise sources such as a distant mixing plant, while the internal monitoring sensor is located deep within the air duct. If the distance difference between the external and internal noise sound waves is approximately 1.02 meters, and the speed of sound is known to be approximately 340 m / s, then theoretically, noise emitted at the same moment will arrive at the internal sensor approximately 3 milliseconds later than at the external sensor. At a sampling rate of 48 kHz, these 3 milliseconds correspond to the displacement of approximately 144 sampling points. To accurately calculate this lag, the algorithm uses a discrete cross-correlation function to measure the similarity between two sequences at different relative displacements. For the current frame data of length N, the cross-correlation calculation formula is defined as... In the formula The time delay index represents a tentative approach; its physical meaning is to shift the external sequence to the right on the time axis. There are several points. This formula calculates by iterating through the internal sequence and summing the point-to-point products of the translated external sequence. When When the value of is exactly equal to the actual physical path delay, the noise waveforms in the two sequences, such as the vibration peaks at a certain instant, will completely overlap, and the sum of the products will reach its maximum value. Therefore, the system determines the optimal delay by searching for the global maximum value of the cross-correlation function. Its mathematical expression is In the example above, the calculation result might be displayed as... The cross-correlation function reaches its extreme value when the value is 144, i.e. =144. After obtaining this parameter, the system performs an alignment operation: aligning the external discrete sequence... Overall index offset Each unit. In practice, this involves data buffer operations, discarding the external sequence before... A new sequence is generated by adding data points and incorporating the tail data from the previous frame or padding with zeros. At this point, the outer sequence has been aligned. Background noise components in the time axis and the internal discrete sequence It achieves strict co-origin and co-phase, meaning that the values ​​of both at the same index n reflect the noise fluctuations at the same physical moment.

[0022] Finally, in step S23, although the signal is time-aligned, the original signal exhibits typical pink noise characteristics in terms of frequency domain energy distribution. This means extremely high low-frequency energy is caused by fan rotation and compressor operation, while high-frequency energy rapidly decays with increasing frequency. However, the micro-icing feature of interest in this application—the airflow disturbance caused by frost crystals altering surface roughness—is mainly manifested as high-frequency turbulent noise above 5kHz. If the original signal is processed directly, the weak frost feature fingerprint in the high-frequency band will be overwhelmed by the huge energy dynamic range in the low-frequency band, leading to a significant decrease in the sensitivity of subsequent feature extraction algorithms. To address this issue, the system introduces pre-emphasis processing, which essentially applies a first-order high-pass filter to flatten the signal spectrum. This filter is sensitive to the rate of change of the signal, suppressing slowly changing low-frequency components while amplifying rapidly changing high-frequency components. The pre-emphasis calculation formula is described in the form of a difference equation: For the internal channel, the output signal... For external channels, the output signal In the formula The pre-emphasis coefficient is determined based on the average spectral attenuation slope of the sound signal, and its value is set between 0.95 and 0.98. Taking 0.97 as an example: the current output value equals the current input value minus 97% of the previous input value. If the signal is low-frequency, meaning the values ​​of adjacent sampling points change very little, The difference When the value is close to 0, low-frequency energy is significantly attenuated; conversely, if the signal is high-frequency, meaning the values ​​of adjacent sampling points change drastically, the difference... A larger value is retained, and high-frequency energy is relatively enhanced. Through this linear transformation, the spectral tilt that decreases with increasing frequency in the original signal is compensated, resulting in a more balanced signal-to-noise ratio across the entire frequency band. The final output is a dual-channel synchronous audio dataset. This dataset is a structured data packet containing two parallel one-dimensional vectors. and These two vectors are precisely synchronized on the time axis (eliminating the sound path difference), and have undergone high-frequency enhancement on the frequency axis (eliminating the spectral tilt), while retaining all the dynamic information of the original airflow sound and background noise.

[0023] In step S3, adaptive background noise cancellation is performed on the dual-channel synchronous audio dataset to obtain a clean airflow friction sound signal. It is understandable that although rigid alignment of the dual-channel signals along the time axis and pre-emphasis compensation in the frequency domain have been completed in previous steps, the external environmental noise actually undergoes a complex physical acoustic modulation process as it penetrates the thick metal chassis of the refrigeration unit and enters the duct. The chassis is not an ideal acoustically transparent body, but rather a complex linear time-varying acoustic filter. It produces transmission attenuation, multipath reflection, and reverberation delay effects on different frequency components of external noise, resulting in a significant difference in waveform between the transmitted sound entering the duct and the original external reference sound. If a simple linear subtraction is used, this nonlinear distortion caused by the transmission path will lead to waveform mismatch. This not only fails to eliminate background noise but may also introduce false high-energy residuals into the differential signal, masking the truly weak early-frosting airflow characteristics. Therefore, it is necessary to construct an adaptive filter that can simulate the acoustic transmission characteristics of the chassis in real time. By learning the evolution of external noise along this physical path, it can accurately predict and reconstruct a copy of the response inside the air duct, thereby achieving precise removal of background noise.

[0024] In one embodiment, step S3 includes: S31: extracting time-series samples from the external reference signal of the dual-channel synchronous audio dataset to construct a reference input vector; S32: using the filter weight vector at the current moment to estimate the noise component of the reference input vector to obtain the estimated background noise component; S33: performing differential processing on the internal mixed signal in the dual-channel synchronous audio dataset and the estimated background noise component to obtain a clean airflow friction sound signal.

[0025] The specific processing is as follows: First, step S31 is executed. In the digital signal processing architecture, this is achieved by constructing a tapped delay line. The filter order is set to M. In the specific sand and concrete cooling scenario, considering the complexity of the chassis structure and the possible reverberation time, in order to cover sufficient acoustic decay periods, M is set to a higher level, for example, M=256. This means that the filter has 256 memory units, capable of backtracking and processing noise history information within the past approximately 5.3 milliseconds at a 48kHz sampling rate. For each discrete time point n, the data flow controller extracts the current time and its data points from the external reference signal channel of the dual-channel synchronous audio dataset, along with the data points from the previous M-1 time points buffered by the delay line. This series of time-correlated scalar values ​​is stacked to construct a vertical column vector, i.e., the reference input vector. Its mathematical construction logic is as follows:

[0026]

[0027] The vector Existing in an M-dimensional Euclidean space, this data not only contains the instantaneous amplitude of the external noise at the current moment but also comprehensively describes the waveform evolution trajectory of the external noise source over a recent period. This data structure allows subsequent algorithms to see not only the present noise but also its past, thus enabling the simulation of causal propagation delays of sound waves within the chassis panel. For example, when processing begins at time n, the first element of the vector is the latest pre-emphasized noise sample obtained from the ADC, while the last element is a sample obtained 255 sampling periods ago. As time n progresses, this vector slides and updates along the time axis, with the latest data moving to the front and the oldest data moving to the back, always maintaining a real-time snapshot of the external noise environment.

[0028] Next, step S32 is executed. In this step, the core component is a set of dynamically adjustable coefficients, called the filter weight vector. This vector is related to the reference input vector. Having the same dimension M, each element contained within it Where k=0,1,…,M-1 each has a clear physical meaning: they represent the transmission response coefficients of the chassis panel to external noise at different delay times. For example, This may represent the attenuation coefficient of the direct transmission component, while This may represent the coefficients of the hysteresis component that reaches the interior after one reflection. At the initial moment, such as when the device is first started, the acoustic characteristics of the chassis are not yet known, and the weight vector... The vector is initialized to either an all-zero vector or a tiny random number vector. During the normal execution cycle of step S32, the reference input vector is weighted and summed using the weight vector that has been iteratively updated at the current time, i.e., a dot product operation. This process is known as Finite Impulse Response (FIR) filtering in signal processing. Its purpose is to synthesize a predicted value using externally acquired historical noise data processed through a virtual transfer function. The calculation formula is: Through this linear combination process, the algorithm is essentially performing a high-dimensional mapping: attempting to fit the background disturbance experienced by the internal air duct at this moment using a linear combination of external reference signals. The resulting scalar result is the estimated background noise component, denoted as . If the weight vector It has converged to the optimal solution, so this It will approximate the actual background interference component transmitted by external noise in the internal mixed signal in terms of waveform, phase, and amplitude. For example, if there is a high-amplitude impact noise, such as the impact of a crusher, this noise appears... In the vector, after passing through the sound insulation characteristics of the chassis. After vector multiplication, the output is the estimated background noise component. It will be a signal with attenuated amplitude and extended waveform, accurately simulating the dull echo of the impact sound entering the air duct.

[0029] Finally, step S33 is executed. This step comprises two parallel but logically tightly coupled processes: signal stripping and model self-learning. First, signal stripping is performed. The internal mixed signals from the dual-channel synchronous audio dataset are stripped... As the desired response, the estimated background noise component generated in step S32 is subtracted from it. The difference at each discrete time point n is obtained. At the signal output level, this is precisely the pure airflow friction sound signal that this method aims to obtain. The logic is based on the fact that the internal mixed signal is a linear superposition of airflow sound and transmitted background sound; and estimating the background noise component is the best estimate of the transmitted background sound. When the estimated background noise component is sufficiently accurate, the subtraction of the two cancels out the background interference, and the remaining residual is naturally the airflow friction sound, independent of external noise. Simultaneously, to ensure the accuracy of the estimated background noise component, the difference must be utilized... The filter itself is modified, specifically by performing a least mean square (LMS) weight update algorithm. The logic here is based on the gradient descent principle: if the difference... The energy is large, indicating that the estimated noise is high. A significant deviation from the actual transmitted background sound indicates that the filter has not yet fully learned the acoustic characteristics of the chassis. Therefore, the weight vector needs to be adjusted according to the magnitude and direction of the error. This minimizes the prediction error at the next time step (i.e., minimizes the mean squared error, MSE). The weight update formula is as follows:

[0030]

[0031] In the formula This is the step size factor, which is also the convergence speed control parameter, and is taken as a very small positive number, such as 0.001. The physical meaning of this formula is: the adjustment of the weights is proportional to the current prediction error. and the current reference input vector The product of. Taking a specific numerical value as an example: such as at time n, the external reference signal... A positive pulse suddenly appears, while the internal mixed signal... A corresponding positive pulse (transmission noise) also appears, but the current filter estimates... Too small, leading to calculation error It is a positive value. According to the update formula, the weight vector The coefficient corresponding to the input at that moment will increase (by adding a positive increment). Therefore, when processing the next time step n+1, if a similar input is encountered, the filter output will... This will increase the value slightly, thus reducing the new error. Through this point-by-point iterative mechanism, the filter weight vector... After thousands of sampling iterations, it will automatically converge and lock onto the true acoustic transfer function of the chassis panel. Even if the external environment changes (such as changes in temperature leading to changes in sound speed, or changes in the location of noise sources), the adaptive algorithm can still continuously... Feedback Dynamic Adjustment As time n continuously flows from 0, 1, 2, ..., a scalar difference will be generated at each moment. . These continuously produced When pieced together in chronological order, a complete and pure airflow friction sound signal is formed.

[0032] In step S4, the pure airflow friction sound signal is subjected to short-time Fourier transform and frequency domain mapping to obtain a high-frequency acoustic signature spectrum. It should be understood that although the pure airflow friction sound signal has been successfully extracted from the aliased sound field after the adaptive background noise cancellation process in the previous stage, this signal still remains in the time domain, appearing as a discrete amplitude sequence flowing linearly with time. In the actual operation of sand and gravel concrete refrigeration equipment, the accumulation process of frost on the evaporator surface is essentially a process of gradual microscopic morphological change leading to the evolution of fluid dynamic characteristics. This evolution is extremely subtle in a simple amplitude-time waveform because changes in the total energy of the airflow are often masked by macroscopic factors such as fan speed fluctuations. Simply observing the peaks and troughs makes it difficult to distinguish whether it is an increase in airflow or a thickening of the frost layer. The nucleation and growth of frost crystals essentially change the surface roughness and boundary layer thickness of the fins. This physical change in microstructure excites extremely small-scale aerodynamic turbulence, thereby generating unique acoustic resonance or modulation effects in specific high-frequency bands. To capture the subtle shifts in frequency components over time—the so-called "soundprint" features—it is necessary to break away from the single time-domain observation perspective and adopt a time-frequency joint analysis method. This method deconstructs the one-dimensional sound wave signal into a three-dimensional energy distribution spectrum. Through short-time Fourier transform and frequency domain mapping, a high-frequency soundprint feature spectrum that can intuitively reflect the evolution of frost formation is constructed.

[0033] In one embodiment, step S4 includes: S41: performing windowing and framing processing on the pure airflow friction sound signal to obtain a windowed frame sequence matrix; S42: performing fast Fourier transform and power spectral density estimation on the windowed frame sequence matrix to obtain a full-band power spectral matrix; S43: locking and extracting the high-frequency aerodynamic feature range of the full-band power spectral matrix to obtain a high-frequency acoustic signature feature spectrum.

[0034] The specific processing is as follows: First, the pure airflow friction sound signal input in step S41 is a non-stationary random process, meaning that its statistical characteristics (such as mean, variance, and spectral distribution) continuously change over time. However, the classic Fourier transform is based on the assumption that the signal is stationary throughout the entire analysis period. If a signal lasting several minutes is directly transformed, all temporal location information will be lost, making it impossible to determine when a certain frequency component appears. To resolve this contradiction, the short-time stationarity assumption is adopted, which assumes that the state of the airflow is relatively stable within extremely short time segments (e.g., tens of milliseconds). Therefore, the slice parameters need to be defined first: frame length... and frame shift With a sampling rate of 48kHz, a frame length of 1024 sampling points was set, corresponding to a physical duration of approximately 21.3 milliseconds. This is sufficient to cover the fundamental frequency period of the airflow sound and ensure local stationarity. To avoid abrupt changes in features between adjacent frames and to fully utilize data edge information, a frame shift of 512 points was set, meaning there is a 50% overlap between adjacent frames. However, simple rectangular slicing (i.e., direct truncation) causes a truncation effect in the time domain, resulting in severe spectral leakage in the frequency domain. This is because the steep edges of the rectangular window in the time domain correspond to infinitely extended sidelobes in the frequency domain as expressed by the Sink function. These high-energy sidelobes leak the energy of strong frequency components into adjacent weak frequency passbands, causing the faint high-frequency characteristics of the frost layer to be masked. Therefore, a smooth window function needs to be multiplied point-by-point on each frame of signal; the Hanning window is chosen here. The Hanning window exhibits a perfect raised cosine bell curve, smoothly compressing both ends of each frame of data to zero, thereby greatly suppressing the sidelobe amplitude. After this processing step, the originally overlapping and continuous one-dimensional long sequence is reorganized into a two-dimensional matrix called the windowed frame sequence matrix. Each column of this matrix represents an independent time frame, with a column length of 1024; as time progresses, the number of columns increases, and each column carries a short acoustic segment that has been smoothed.

[0035] Next, step S42 is executed. The core task of this step is to transform the dimension of the data description from amplitude-time to energy-frequency. For each column in the windowed frame sequence matrix (i.e., each frame of data), the Fast Fourier Transform (FFT) algorithm is applied. The FFT is an efficient implementation of the Discrete Fourier Transform (DFT), which maps a time-domain vector containing 1024 real-valued samples to a frequency-domain vector containing 1024 complex elements. Mathematically, for the... The first frame The complex spectrum of a frequency point The calculation follows the STFT formula: Where k is the frequency index, with a value ranging from 0 to... -1, The number of points for the FFT is equal to the frame length, which is 1024. For the windowed frame sequence matrix, the first... The data in the frame. At this point, the resolution in the frequency domain... Determined by the sampling rate and the number of FFT points, the calculation yields... =48000 / 1024 = 46.875Hz. This means that index k=1 represents 46.875Hz, index k=10 represents 468.75Hz, and so on. It is a complex number containing a real part and an imaginary part, corresponding to amplitude and phase information, respectively. In the scenario of frost monitoring, the absolute value of the phase is often significantly affected by the randomness of the signal's initiation time, and is therefore not very meaningful. What truly carries information about the intensity of airflow turbulence is the energy magnitude in each frequency band. Therefore, it is necessary to further calculate the power spectral density. By taking the square of the complex modulus, the complex spectrum is converted into a real power spectrum: At this time, it is generated It is a full-band power spectrum matrix. The rows of this matrix correspond to the frequency axis (extending from the 0Hz DC component to the 24kHz Nyquist frequency), and the columns correspond to the time axis (frame index). Each element in the matrix... The value intuitively quantifies the value at the first... Within a time segment, the frequency is How much energy does the sound component possess?

[0036] Finally, in step S43, the frosting process on the evaporator surface begins with the nucleation of microscopic ice crystals. These tiny ice crystals are extremely small (micrometer to millimeter in size) and adhere to the smooth metal fins, effectively increasing the roughness of the surface. When high-speed cold air flows over these rough surfaces, it induces microscale shedding eddies. According to acoustic principles, the smaller the scale of the eddies, the higher the frequency of the radiated noise. Therefore, the acoustic manifestation of early frosting is concentrated in the high-frequency broadband noise enhancement, exhibiting a sharp hissing sound characteristic. In contrast, low-frequency sounds are more associated with large-scale structural vibrations. To focus on this core characteristic, the system presets a high-frequency fingerprint range, for example, by setting a starting frequency. =5000Hz and cutoff frequency =15000Hz. First, these physical frequency boundaries need to be mapped to row indices in the FFT matrix:

[0037]

[0038]

[0039] This means that only rows 107 to 320 of the full-band power spectrum matrix truly contain frosting information. Subsequently, matrix slicing and reconstruction operations are performed to construct a new feature matrix. It only retains the full-band matrix arrive The lines between them. The formula for high-frequency fingerprint interval extraction is expressed as: ,in In this new matrix, row index Starting from 0, corresponding to 5000Hz in physical frequency, the last row corresponds to 15000Hz. After this elimination operation, the semi-spectral matrix, which could have had as many as 512 rows, is reduced to a dedicated spectrum of approximately 213 rows. The final high-frequency acoustic signature feature spectrum is obtained. It is a highly condensed acoustic profile that shields against low-frequency mechanical noise interference and removes potential electronic thermal noise in the ultra-high frequency range (>15kHz), purely displaying the aerodynamic noise energy distribution directly related to the microscopic roughness of the fin surface. Each data column is like a slice specimen, recording the specific high-frequency energy fingerprint created by the airflow friction on the surface of the microscopic ice crystals at that moment.

[0040] In step S5, the acoustic fingerprint of micro-roughness is quantitatively extracted from the high-frequency acoustic fingerprint feature map to obtain the acoustic roughness index. Correspondingly, although the subtle shift in energy distribution with the frost formation process can be visually observed from the high-frequency acoustic fingerprint feature map, this type of data is still too abstract and involves a large amount of data for industrial control systems. To achieve precise automated monitoring and control decisions, it is necessary to compress, abstract, and quantify this complex two-dimensional frequency domain image information into a one-dimensional numerical index that can directly reflect the physical state of the evaporator surface. Physics research shows that the accumulation of frost on the fin surface has a dual acoustic effect: on the one hand, the random micro-turbulence induced by the rough surface causes the spectral structure of airflow noise to change from ordered to disordered, manifested as an increase in spectral entropy; on the other hand, the increase in turbulence intensity leads to an increase in the overall sound pressure level in the high-frequency band. Therefore, by comprehensively considering the disorder and energy intensity of the spectral morphology, and performing multi-dimensional mathematical analysis and fusion calculations on the high-frequency acoustic fingerprint feature spectrum, it is possible to quantitatively extract the micro-roughness acoustic fingerprint that characterizes the changes in micro-roughness, and finally generate a stable, monotonous acoustic roughness index with clear physical meaning.

[0041] Figure 3 This is a flowchart of step S5 in the intelligent monitoring method for a sand and gravel concrete cooling equipment according to an embodiment of this application. Figure 3As shown, in one embodiment, step S5 includes: S51: normalizing the probability density of the high-frequency acoustic signature feature spectrum to obtain a normalized spectral distribution matrix; S52: estimating the frequency disorder of the normalized spectral distribution matrix to obtain a spectral entropy feature vector; S53: calculating the average power intensity of the high-frequency acoustic signature feature spectrum within the feature frequency band to obtain an in-band energy feature vector; S54: performing weighted fusion and moving average filtering on the spectral entropy feature vector and the in-band energy feature vector to obtain an acoustic roughness index.

[0042] The specific processing is as follows: First, step S51 is executed. The input high-frequency voiceprint feature spectrum is a two-dimensional matrix, where rows correspond to frequencies and columns correspond to times. For each column in the matrix, i.e., each time frame... The algorithm first iterates through all frequency points in the column. The power spectrum values ​​are used to calculate the total energy value of the frame. Then, a pointwise division operation is performed, dividing the power value of each frequency point in the column by the total energy value. After this operation, the data in this column is transformed into a new probability vector whose sum of all elements is strictly equal to 1. This essentially treats the spectrum of each frame as a discrete probability distribution function, where each frequency point is considered a value of a random variable, and its corresponding value represents the probability of that frequency occurring. Repeating this operation for all time frames creates the normalized spectral distribution matrix, denoted as . .

[0043] Next, step S52 introduces Shannon's information entropy theory to quantify the disorder of airflow noise. Physically, when the evaporator fin surface is clean and smooth, the airflow passing over the fins mainly produces wind shearing sounds or vortex shedding sounds at specific frequencies, with energy concentrated on a few characteristic frequencies. At this time, the spectral distribution exhibits a peak-like shape, corresponding to low probability distribution uncertainty, i.e., a low entropy value. However, when microscopic frost crystals begin to nucleate and accumulate on the surface, the surface becomes rough and uneven, inducing a large amount of random microscale turbulence, causing the airflow noise to evolve into broadband white noise. The energy is evenly distributed throughout the high-frequency band, at which point the spectral distribution tends to be flat, with maximum uncertainty, i.e., a high entropy value. The entropy value is calculated for each column of the normalized spectral distribution matrix: , in the formula, For each column of the normalized spectral distribution matrix, The width of the characteristic frequency band (i.e., the number of frequency points). It is a very small positive number, such as This is used to prevent the mathematical error of log(0) in logarithmic operations. Calculation result It is a scalar value, which changes with time frames. The advancement of this process forms a one-dimensional time series, namely the spectral entropy feature vector. For example, in the early stages of frost formation, the value of this vector may fluctuate around 4.5; as the frost layer thickens, the value gradually climbs to above 6.0, sensitively capturing the whitening trend of the spectral structure.

[0044] Then, step S53 is executed. Although spectral entropy can sensitively reflect changes in waveform structure, relying solely on entropy values ​​may overlook absolute changes in energy intensity. In fact, the turbulence caused by frost not only flattens the spectrum but also directly leads to a significant increase in the overall sound pressure level in the high-frequency band, i.e., a louder hissing sound. To capture this feature, the average power of the high-frequency acoustic signature feature spectrum within the feature frequency band is calculated and converted to a logarithmic decibel scale. For each frame... The formula for calculating the energy within the band is:

[0045]

[0046] The formula first calculates the arithmetic mean of the power values ​​of all characteristic frequency points within the frame to obtain the average power density of that frequency band. Then, it takes the logarithm to base 10 and multiplies by 10 to convert it to a decibel value. The calculated... This constitutes the in-band energy feature vector. This vector intuitively reflects the loudness changes of high-frequency noise. For example, in a clean state, the in-band energy may remain at -60dB; while in severe frosting, due to intense airflow friction, this value may rise to -45dB. This feature, as a powerful complement to the spectral entropy, ensures that the system can still detect anomalies even when the spectral shape does not change significantly but the noise increases.

[0047] Finally, step S54 is executed. Since spectral entropy and energy are two physical quantities with completely different dimensions (one is a dimensionless information quantity, and the other is a decibel value), and their sensitivity to the frosting state may differ at different stages, they need to be merged into a unified dimensionless index through weighted fusion. The fusion calculation first involves normalization. An empirical maximum value is set. For example, the theoretical maximum entropy is 8.0 and For example, the absolute value of the system's maximum design noise -20dB is mapped to the interval [0,1]. Then, a linear weighted summation formula is applied: In this formula, and These are preset weighting coefficients, and they add up to 1. The setting of these two coefficients is based on calibration using a large amount of experimental data. Generally, given that spectral entropy is more sensitive to early microscopic frost formation (frost crystal nucleation stage), while energy characteristics perform better in the middle and later stages of frost formation (ice thickening stage), to emphasize early warning capabilities, a higher weight is tended to be assigned to spectral entropy, for example, setting... =0.6, =0.4. Preliminary fusion index obtained. Although it integrates information from two dimensions, the randomness of airflow can cause severe sawtooth fluctuations in the sequence over short periods, which is detrimental to determining stable states. Therefore, time-domain smoothing is necessary. A moving average filter is used, with a set smoothing window length. For example, the number of data frames corresponding to 5 seconds. The final acoustic roughness index. The formula is as follows: This smoothing process filters out transient high-frequency disturbances while preserving the steady-state trend reflecting the evolution of the frosting state. The final output acoustic roughness index... It is a numerically smooth, monotonically increasing time series, whose value gradually climbs from 0.2 to 0.8, directly and linearly representing the current state of the micro-roughness of the evaporator surface.

[0048] In a preferred embodiment, step S5 includes: S5-1: Normalizing the probability density of the high-frequency acoustic signature feature spectrum to obtain a normalized spectral distribution matrix; S5-2: Calculating the Renyi generalized spectral entropy based on Strouhal number weighting on the normalized spectral distribution matrix to obtain a spectral entropy feature vector; S5-3: Calculating the average power intensity of the high-frequency acoustic signature feature spectrum within the feature frequency band to obtain an in-band energy feature vector; S5-4: Performing weighted fusion and moving average filtering on the spectral entropy feature vector and the in-band energy feature vector to obtain an acoustic roughness index. Since the implementations of S5-1, S5-3, and S5-4 have been explained in S51, S53, and S54 above, they will not be repeated here, but the focus will be on the specific implementation of S5-2.

[0049] It is understandable that, although Shannon entropy was mentioned as a metric in the aforementioned basic scheme, in high-intensity sand and gravel refrigeration industrial environments, background noise typically exhibits a steady-state broadband noise distribution close to Gaussian, while turbulent noise caused by frost layers has typical non-Gaussian distribution characteristics. Since Shannon entropy mathematically measures the average uncertainty of a probability distribution, it assigns equal weight to changes in probability distribution across all frequency bands. This leads to Shannon entropy easily saturating in low signal-to-noise ratio environments, making it difficult to effectively distinguish between simple background white noise enhancement and complex turbulent chaotic signals caused by microscopic roughness. Furthermore, physical fluid dynamics mechanisms indicate that higher frequencies correspond to smaller eddy scales and greater sensitivity to early frost formation on microscopic surfaces. To overcome the limitations of traditional statistical methods and enhance the system's ability to detect micrometer-level early frost in high-noise environments, this embodiment employs a Renyi generalized spectral entropy (S-WRSE) calculation model based on Strouhal number weighting. This model combines prior physical knowledge with generalized statistical theory, and through nonlinear weighting, accurately transforms unstructured voiceprint information into structured physical indicators that are highly sensitive to microscopic frost conditions.

[0050] Based on this, in one embodiment, step S5-2: calculating the Renyi generalized spectral entropy based on the Strouhal number weighting of the normalized spectral distribution matrix to obtain the spectral entropy eigenvector includes:

[0051] Weights are applied to each frequency point in the normalized spectral distribution matrix to construct a frequency-sensitive weight vector. This step aims to introduce a fluid dynamics weighting mechanism, physically enhancing the algorithm's responsiveness to microscopic roughness. Based on the Strauhal number principle in fluid dynamics, the frequency of vortex shedding generated by fluid flowing over obstacles (i.e., frost crystals) is... Characteristic dimensions of the obstacle Inversely proportional ( In the early stages of micro-frost formation, the crystal nuclei are extremely small (micrometer-scale), and the turbulent energy they generate is mainly concentrated in the extremely high frequency range. To match this physical characteristic, the algorithm constructs a frequency-sensitive weight vector that exhibits an exponentially increasing trend. For the first characteristic frequency band in the normalized spectral distribution matrix... Each frequency point has its corresponding fluid dynamics sensitivity weight. The calculation formula is constructed as follows:

[0052]

[0053] In the formula, Indicates the first The actual physical frequency (unit: Hz) corresponding to each frequency point; It is the geometric center frequency of the characteristic frequency band, which serves as the reference point for normalization; The preset sensitivity coefficient was determined by simulating the vortex shedding frequency response characteristics of frost crystals of different sizes in a laboratory environment, with the goal of maximizing the spectral entropy distinction between the initial frosting stage and the clean state. When set... When the value is greater than 1, the function exhibits nonlinear amplification of high-frequency components, thereby suppressing the contribution of low-frequency components to the entropy value. Specifically, if the characteristic frequency band ranges from 5kHz to 15kHz, then the center frequency... Set to 10kHz, which is 10000Hz. Preset sensitivity coefficient. =1.5 to appropriately amplify the high-frequency contribution. If the current calculated frequency point... corresponding physical frequency =15000Hz, then its weight is calculated as follows =(1+15000 / 10000)^{1.5}=3.95; while for lower frequencies... =5000Hz, its weight is only (1+0.5)^{1.5}=1.83. This difference in weight allocation ensures that the algorithm mainly focuses on those ultra-high frequency turbulence features excited by tiny frost crystals.

[0054] Based on the frequency-sensitive weight vector, the normalized spectral distribution matrix is ​​weighted and aggregated using Renyi probability moments to obtain a weighted Renyi generalized spectral entropy vector. This step uses Renyi generalized entropy to replace Shannon entropy, and the core lies in introducing an order. To adjust the sensitivity to the shape of the probability distribution. When When the value is greater than 1, the algorithm no longer simply sums the peak values ​​(i.e., the main turbulence components) in the probability distribution, but instead performs exponential amplification. This allows the algorithm to effectively suppress flat and low-amplitude background noise, thereby significantly improving the signal-to-noise ratio gain for non-Gaussian burst signal characteristics. For each time frame... Weighted Renyi generalized spectral entropy The calculation formula is expressed as follows:

[0055]

[0056] In the formula, It is the normalized probability density value output in step S51, that is, each column in the normalized spectral distribution matrix; It is the order of Renyi entropy, preferably in the range of [2,3], such as taking 3 to enhance the capture of the main turbulent components); This represents the total bandwidth and number of frequency points of the characteristic frequency band. The numerator of the formula calculates the weighted probability power moment, while the denominator is the sum of the weights, used to maintain mathematical normalization. Continuing the example above: Set =3, for a certain high-frequency noise point, such as its normalized probability density =0.1, its contribution in Shannon entropy is 0.1×log(0.1), while in Renyi entropy its core term becomes 0.1^3=0.001. This exponential operation greatly widens the numerical gap between strong signals (frost signals) and weak signals (background noise).

[0057] The weighted Renyi generalized spectral entropy vector is amplitude-normalized to obtain the spectral entropy eigenvector. Considering that Renyi entropy calculation involves complex logarithmic and exponential operations, its original output value may contain negative values ​​or fluctuate significantly over its dynamic range. To ensure that the output index can intuitively and stably reflect the thermo-acoustic complexity of the system, its magnitude needs to be extracted. The calculation formula is as follows: After this processing, the resulting S-WRSE spectral entropy feature vector has a larger dynamic range and higher sensitivity compared to the traditional entropy value. Since the signal-to-noise ratio of this index has been significantly improved through physical weighting and statistical amplification, in the subsequent fusion calculation in step S5-4, the weight of the in-band energy feature can be appropriately reduced, and more reliance can be placed on this high-confidence morphological index to accurately identify early micron-level frost.

[0058] In step S6, based on a preset system cleaning baseline fingerprint, the acoustic roughness index is intelligently judged based on the evolution of acoustic fingerprints to determine the current frosting phase state of the evaporator. It is understandable that the acoustic roughness index, as a monotonically changing numerical sequence, objectively records the evolution trajectory of the microscopic state of the evaporator surface. However, in actual industrial control scenarios, a single absolute value often lacks universal judgment effectiveness. This is because the manufacturing tolerances, installation environment, and real-time operating conditions such as fan speed settings and initial aggregate temperature vary for each refrigeration unit, resulting in different initial acoustic roughness reading noise levels under completely clean conditions. Simply setting a fixed absolute threshold (e.g., an alarm when the index is greater than 0.5) can easily lead to frequent false alarms or missed alarms due to individual equipment differences or zero-point drift of sensors after long-term operation. Furthermore, frosting is a continuous nonlinear gradual process from nothing to something, from microscopic crystal nuclei to macroscopic ice layers. Its state boundaries have inherent ambiguity, and there is no clear physical boundary to distinguish between normal and micro-frost. To this end, this application introduces a reference frame based on relative change rate and fuzzy logic reasoning method. Based on a preset system cleaning benchmark fingerprint, the acoustic roughness index is intelligently judged based on the frosting state of the acoustic fingerprint evolution to determine the current frosting phase state of the evaporator.

[0059] In one embodiment, step S6 includes: S61: calculating the relative rate of change of the acoustic roughness index relative to the system cleaning reference fingerprint to obtain the feature offset; S62: performing multi-phase fuzzy membership evaluation on the feature offset to obtain the fuzzy membership vector; S63: performing maximum membership decision and time-domain de-jitter latching on the fuzzy membership vector to obtain the frosting phase state.

[0060] The specific processing is as follows: First, step S61 is executed. To overcome the limitations of fixed threshold judgment, this method adopts a dynamic benchmark relative comparison method. The core lies in establishing a dynamically updated reference scale, namely the system cleaning benchmark fingerprint. This fingerprint is a dynamic variable stored in the controller's non-volatile memory (such as EEPROM). Its acquisition mechanism is based on a self-learning logic: whenever the refrigeration unit completes a complete defrosting cycle and confirms the resumption of refrigeration operation, such as the 5th to 10th minute after startup, the evaporator surface is theoretically in its cleanest and frost-free state. The system automatically collects the acoustic roughness index within this time window, calculates its arithmetic mean, and locks it as the current system cleaning benchmark fingerprint, denoted as: For example, the baseline fingerprint measured after a defrost cycle is 0.25. This mechanism ensures that each monitoring start point is based on the latest physical state of the current equipment, effectively offsetting baseline drift caused by dust accumulation on the wind turbine blades or sensor aging. During real-time monitoring, the acoustic roughness index input at each time t is denoted as... The algorithm no longer focuses on its absolute size, but instead calculates its percentage degradation relative to the baseline fingerprint, i.e., the feature offset. The calculation logic is as follows: If the currently monitored roughness index rises to 0.30, then the feature offset... =(0.30-0.25) / 0.25×100=20. The resulting value is a dimensionless percentage, intuitively representing that the surface roughness of the evaporator has deteriorated by 20% compared to its perfect state. This approach allows refrigeration units of different power and models to be managed uniformly using the same evaluation standard, such as a 20% offset, greatly improving the algorithm's generalization ability.

[0061] Following step S62, since the transition from clean to severe frost is a smooth, gradual process, and acoustic characteristics often fluctuate in the transition zone, traditional binary logic (black and white) is insufficient to accurately describe this intermediate state. Therefore, this step introduces fuzzy set theory, defining three core phase spaces: 1. Clean Zone: Represents the evaporator operating at high efficiency with no significant frost accumulation. 2. Crystal Nucleation Warning Zone: This is the most crucial early stage, indicating that frost crystals have formed in large numbers at the microscopic level, significantly increasing the surface roughness of the fins, but before forming macroscopic blockages. This is the optimal time for optimization and calibration. 3. Growth Blockage Zone: Represents the frost layer having grown to a certain thickness, beginning to physically block the air duct, requiring immediate and forceful measures. The algorithm utilizes a preset fuzzy membership function to offset the scalar feature... Mapped to a three-dimensional vector containing three elements. These three elements represent the probability or degree of membership of the current state to the three sets mentioned above, and their values ​​range from [0,1]. The specific calculation depends on the designed piecewise linear membership function. For the membership degree of the clean zone... Using a left-shoulder trapezoidal function:

[0062]

[0063] This means that when the degradation level is less than 5%, the system is 100% certain of being in a clean state; when the degradation level is between 5% and 10%, the certainty decreases linearly; and when it exceeds 10%, it is considered no longer in a clean state. (Regarding the membership degree of the crystal nucleus early warning zone...) Using triangular or trapezoidal functions:

[0064]

[0065] The function peaks at 1 when the offset is between 10% and 20%, with the intervals [5,10] and [20,30] representing the transition region. This accurately defines the quantization range for early frosting. (Membership to the growth blockage region is also discussed.) Using a right-shoulder trapezoidal function:

[0066]

[0067] When the offset exceeds 30%, the system is certain it has entered a severe blocking phase. (Based on the previously calculated offset...) Let's take 20 as an example and substitute it into the calculation: =0, because 20 > 10; =1, which falls exactly in the interval [10,20]; =(20-20) / 10=0, at this time the fuzzy membership vector is: =[0,1,0], indicating that the system is in a typical crystal nucleus early warning state. If =25, then the calculation result is =0, =0.5, =0.5 indicates that the system is in a critical transition period from early warning to worsening congestion.

[0068] Finally, step S63 is executed. Although the fuzzy vector provides detailed probabilistic information, the actuator (controller) requires a unique and definite status code to perform the operation. First, based on the principle of maximum membership, the system compares... The state corresponding to the component with the largest value among the three components is selected as the instantaneous state. In the above In the example where =25, if the following occurs =0.5 and A tie of 0.5 is, based on the principle of risk priority, classified as a more severe stage, i.e., the instantaneous state is determined to be a growth blockage zone. However, instantaneous electromagnetic interference or sudden airflow changes in the field environment can cause millisecond-level jumps in the acoustic index, thus... Short-term fluctuations, such as repeated switching between cleaning and warning states, can occur. To eliminate the impact of these false signals on the control logic and prevent frequent operation of actuators (such as valves and fan frequency converters), a time-domain debouncing latch mechanism is introduced. This mechanism is implemented using a sliding time window counter. A confirmation time threshold is set. The sampling period is 10 seconds. If the sampling period is 1 second, the corresponding time window length T = 10. The algorithm continuously monitors the latest instantaneous state flow. Only when a state remains unchanged for the past T consecutive sampling moments will it be recognized as the final output state. The logical description formula is:

[0069]

[0070] In the formula, This is an indicator function, taking the value 1 when the condition is true and 0 otherwise. The summation term counts the number of times the instantaneous state equals the current instantaneous state within the current and the past T-1 time intervals. Only when this count is strictly equal to the window length T, i.e., the same state is determined for 10 consecutive seconds (e.g., a crystal core warning is determined for 10 consecutive seconds), will the system update the global state latch. Switch to the new state. Otherwise, the system will remain in the state from the previous moment. The inertial update mechanism greatly improves the system's anti-interference capability, ensuring that the final output frosting phase state is a stable, reliable control signal with high confidence in trend changes.

[0071] In step S7, an optimized calibration control command is generated based on the frosting phase state. This command is then sent to the refrigeration unit controller to execute proactive defensive adjustments. In other words, in the preceding steps, through multi-dimensional acoustic feature analysis and fuzzy logic decision-making, the intelligent monitoring scheme has successfully locked the current frosting phase state of the evaporator, quantifying the previously invisible microscopic physical process into discrete digital status codes. However, simply achieving state perception and diagnosis is insufficient to solve the energy efficiency degradation problem caused by frosting in a closed loop. Without immediate intervention, monitoring data ultimately becomes merely a record for post-fault analysis. Traditional refrigeration control logic often operates in a passive waiting state until the frost thickness reaches a macroscopic threshold to trigger the defrosting procedure. This forces the equipment to undergo an inefficient cycle of normal operation—energy efficiency degradation—shutdown for defrosting—restart. To break this deadlock and achieve true feedforward active defense, this application generates optimized calibration control commands based on the frost phase state and sends them to the refrigeration unit controller. By utilizing the rapid response characteristics of hardware such as variable frequency fans and electronic expansion valves, targeted aerodynamic or thermodynamic adjustments are implemented in the early stage of frost crystal nucleation or the middle stage of growth, thereby suppressing the evolution rate of the frost layer and extending the unit's efficient operating time window.

[0072] In one embodiment, step S7 includes: S71: performing logical branch mapping on the value of the frosting phase state to determine the corresponding strategy mode identifier; S72: obtaining the actuator target parameter vector according to the strategy mode identifier; S73: mapping the physical operating parameters in the actuator target parameter vector to hardware register values ​​and attaching a cyclic redundancy check code to obtain the optimized calibration control instruction.

[0073] The specific processing is as follows: First, step S71 is executed. The input to this step is the stable frosting phase status code output by the previous module after debouncing and latching, with values ​​of 0, 1, and 2. The core of the algorithm is a predefined strategy lookup table, which establishes a unique mapping relationship between the status code and the control strategy mode. For different detected phases, the system defines three standard defensive control modes: The first is the hold / cruise mode (Mode 0), corresponding to an input state of 0 (clean state). In this mode, the evaporator surface is in an ideal heat exchange condition, and the control logic determines that no additional disturbance adjustment is needed. The current PID control parameters should remain unchanged to ensure that the unit operates at the optimal energy efficiency ratio at the predetermined setpoint. The second is the pneumatic stripping mode (Mode 1), corresponding to an input state of 1 (crystal nucleation warning state). Physical principles show that in the early stage of frost crystal nucleation, its adhesion to the fin substrate mainly comes from weak van der Waals forces and hydrogen bonding, and the crystal structure is fragile. At this point, if an unsteady airflow disturbance is introduced, the resulting shear stress is sufficient to destroy the unsolidified crystal nuclei and peel them off the surface. Therefore, this mode aims to activate the dynamic response function of the fan system. The third mode is the thermal suppression mode (Mode 2), corresponding to input state 2 (growth blockage state). When the frost layer has begun to grow and cover the fins, the aerodynamic peeling effect weakens. At this time, the strategy shifts to thermodynamic regulation, by fine-tuning the evaporation temperature of the circulating working fluid to approach but not exceed the freezing point of water, such as raising it to around -2℃, thereby reducing the supercooling of the frost layer surface and slowing down the sublimation rate of water vapor. In other words, by raising the temperature to suppress frost, the irreversible frosting process at this time is delayed, allowing for a longer production time. The logic determination unit quickly retrieves and locks the corresponding strategy code based on the value of the current input frosting phase state and outputs the strategy mode identifier. For example, if the input value for the frosting phase state is 1, then... =1.

[0074] Next, step S72 is executed to construct an action vector containing two physical dimensions. ,in The target frequency unit for the wind turbine is Hz. The target opening degree of the bypass valve is represented by a percentage. When =0 (cruise mode): The system adopts a laissez-faire strategy, directly reading the current setpoint of the main control loop. For example, if the rated frequency under the current load demand is 50Hz, the bypass valve is fully closed to ensure maximum cooling capacity. Then the output vector is: =[50.0,0.0]. When When =1 (pneumatic stripping mode): the control logic no longer outputs a constant fan frequency, but instead outputs the rated base frequency. A low-frequency sinusoidal modulated signal, which takes into account the inertial constraints of the mechanical system, is superimposed on top to generate pulsating airflow. The calculation formula is as follows: Considering the bearing load capacity of high-power industrial fans and the physical limitations of the acceleration / deceleration ramp time of frequency converters, a smoothed and optimized modulation amplitude is set here. =3Hz, and a lower oscillation frequency =0.1Hz, meaning a smooth frequency increase / decrease cycle is completed every 10 seconds. This implies that the fan speed will slowly and continuously fluctuate periodically between 47Hz and 53Hz. This unsteady speed fluctuation will generate alternating pulsating wind pressure in the fin gaps, thereby inducing changes in the microscopic aerodynamic shear stress on the surface and disrupting the adsorption equilibrium of unstable crystal nuclei. Meanwhile, the bypass valve remains closed, i.e. =0, to maintain the low-temperature environment of the evaporator and utilize the brittle effect to assist in stripping. At this time, the output action vector is not a static value, but a function sequence that changes with time t, or, at the instruction level, a parameter package to enable oscillation mode. When When the frost layer is at setting =2 (thermal suppression mode): the primary concern is excessively rapid frost growth. To suppress this growth without completely stopping cooling, the algorithm calculates an optimal hot gas bypass opening, directly injecting a portion of the high-temperature, high-pressure gas discharged from the compressor into the evaporator inlet to artificially increase the evaporation pressure. The goal is to reduce the surface temperature of the evaporator coils. The temperature is controlled within a sensitive range slightly below freezing, such as -3°C to -1°C. The calculation logic is based on a pre-defined feedforward PID model or a lookup table method. , This is the proportional gain coefficient, which is adjusted through on-site testing. The target evaporator surface temperature, This represents the current measured surface temperature of the evaporator. The base feedforward opening is a preset empirical value, representing the approximate valve opening required to raise the temperature to the target range under standard operating conditions. For example, if the preset value is 10%, and calculations show that 15% bypass is needed to neutralize the excessively low evaporation temperature, while the fan maintains a constant 50Hz operation to ensure basic cooling output, then the output vector is: =[50.0,15.0].

[0075] Finally, step S73 is executed. Low-level communication in industrial control environments is typically based on Modbus RTU or CAN bus protocols. These protocols do not directly recognize floating-point physical quantities like 50.0Hz or 15.0%, but rather need to map them to 16-bit or 32-bit integer register values ​​at specific addresses. First, quantization mapping is performed. For example, the frequency control register (address 0x1001) of a fan inverter has a resolution of 0.01Hz, with a range of 0-6000 corresponding to 0-60Hz; the opening control register (address 0x2005) of an electronic valve has 12-bit precision, with a range of 0-4095 corresponding to 0-100%. For the vector [50.0, 15.0] in Mode 2: the fan register value = 50.0 × 100 = 5000 (decimal), which is 0x1388 in hexadecimal. The valve register value = 15.0 / 100.0 × 4095 = 614 (decimal), which is 0x0266 in hexadecimal. Next, the instruction frame is assembled. A standard control instruction frame structure includes: [Device Address] + [Function Code] + [Register Start Address] + [Data Length] + [Data Payload] + [Checksum]. To ensure the security of data transmission in environments with strong electromagnetic interference, a Cyclic Redundancy Check (CRC-16) code must be calculated. The algorithm uses a standard polynomial... A modulo-2 division operation is performed on all preceding bytes, and the resulting 2-byte checksum is appended to the end of the instruction frame. The final optimized calibration control instruction may be a stream of hexadecimal bytes, such as: 01 10 10 01 00 02 04 13 88 02 66 [CRC_L] [CRC_H]. This instruction is immediately sent to the PLC main controller of the refrigeration unit via the RS485 bus. After the controller parses and verifies the instruction, it directly drives the underlying hardware: the fan immediately maintains 50Hz, and the bypass valve driver steps to position 614, thus achieving proactive defensive regulation of the evaporator frosting process at the physical level through precise actuator action.

[0076] In summary, an intelligent monitoring method for sand and gravel concrete refrigeration equipment based on the embodiments of this application has been clarified, addressing the technical pain point of difficulty in early detection of evaporator frost formation in sand and gravel concrete refrigeration equipment under high dust and high humidity environments. This solution does not rely on easily contaminated contact probes or lag-prone differential pressure switches, but instead utilizes dual-channel sensors deployed in parallel inside and outside the air duct to collect mixed signals including airflow sound and ambient background noise. Through timestamp synchronization and adaptive background noise cancellation processing, the pure airflow friction sound signal is effectively extracted. Given that the nucleation and accumulation of frost crystals on the evaporator surface microscopically alters the physical roughness of the fins, thereby specifically modulating the high-frequency acoustic characteristics of the flowing air, the acoustic roughness index is calculated by quantifying and extracting the microscopic roughness acoustic fingerprint from the high-frequency acoustic fingerprint feature spectrum. Based on this, the system monitors the acoustic fingerprint evolution trajectory, accurately determines the current frost phase state, and generates optimized calibration control commands at the initial stage of frost formation, achieving proactive defensive adjustment of the refrigeration unit, thus solving the problems of monitoring failure and lag-prone defrosting control in existing technologies.

[0077] Figure 4 This is a block diagram of an intelligent monitoring system for a sand and gravel concrete cooling equipment according to an embodiment of this application. Figure 4 As shown, the intelligent monitoring system 100 for sand and gravel concrete refrigeration equipment according to an embodiment of this application includes: a signal acquisition module 110, used to acquire an internal mixed signal containing airflow sound and background transmitted sound, and an external reference signal containing only environmental background noise, through monitoring channel sensors deployed in parallel within the refrigeration unit's air duct and a reference channel sensor deployed outside the unit's chassis; a signal synchronization module 120, used to perform time-stamp synchronization alignment and pre-emphasis processing on the internal mixed signal and the external reference signal to obtain a dual-channel synchronous audio dataset; a background noise cancellation module 130, used to perform adaptive background noise cancellation on the dual-channel synchronous audio dataset to obtain a pure airflow friction sound signal; and high-frequency acoustic signature features. Extraction module 140 is used to perform short-time Fourier transform and frequency domain mapping processing on the pure airflow friction sound signal to obtain a high-frequency acoustic fingerprint feature spectrum; acoustic roughness analysis module 150 is used to perform quantitative extraction of micro-roughness acoustic fingerprint from the high-frequency acoustic fingerprint feature spectrum to obtain an acoustic roughness index; frosting state discrimination module 160 is used to perform intelligent frosting state discrimination based on acoustic fingerprint evolution on the acoustic roughness index to determine the current frosting phase state of the evaporator; control command generation and distribution module 170 is used to generate optimized calibration control commands based on the frosting phase state, and the optimized calibration control commands are distributed to the refrigeration unit controller to perform active defensive adjustment.

[0078] Here, those skilled in the art will understand that the specific operations of each step in the intelligent monitoring system for the aforementioned sand and gravel concrete cooling equipment have been referenced above. Figures 1 to 3The intelligent monitoring method for sand and gravel concrete refrigeration equipment is described in detail in the description of the method, and therefore, its repeated description will be omitted.

Claims

1. A method for intelligent monitoring of sand and gravel concrete refrigeration equipment, characterized in that, include: S1: By using monitoring channel sensors deployed in parallel within the chiller unit's air duct and reference channel sensors deployed outside the unit's chassis, internal mixed signals containing airflow noise and background transmitted noise, as well as external reference signals containing only ambient background noise, are collected. S2: Perform time-stamp synchronization alignment and pre-emphasis processing on the internal mixed signal and the external reference signal to obtain a dual-channel synchronized audio dataset; S3: Perform adaptive background noise cancellation on the dual-channel synchronous audio dataset to obtain a clean airflow friction sound signal; S4: Perform short-time Fourier transform and frequency domain mapping on the pure airflow friction sound signal to obtain a high-frequency acoustic signature spectrum. S5: Quantitatively extract the acoustic fingerprint of micro-roughness from the high-frequency acoustic fingerprint feature spectrum to obtain the acoustic roughness index. S6: Based on the preset system cleaning benchmark fingerprint, the acoustic roughness index is intelligently judged based on the frosting state according to the acoustic fingerprint evolution to determine the current frosting phase state of the evaporator; S7: Based on the frosting phase state, generate an optimized calibration control command, which is then sent to the refrigeration unit controller to perform active defensive adjustments.

2. The intelligent monitoring method for sand and gravel concrete refrigeration equipment according to claim 1, characterized in that, Step S2 includes: S21: Perform parallel discretization and quantization on the internal mixed signal and the external reference signal to obtain the internal discrete sequence and the external discrete sequence; S22: Perform time delay estimation and alignment of the internal discrete sequence and the external discrete sequence based on the cross-correlation function to obtain the aligned external sequence; S23: Perform spectral tilt compensation on the internal discrete sequence and the aligned external sequence to obtain a dual-channel synchronized audio dataset.

3. The intelligent monitoring method for sand and gravel concrete refrigeration equipment according to claim 1, characterized in that, Step S3 includes: S31: Extract timing samples from the external reference signal of the dual-channel synchronous audio dataset to construct a reference input vector; S32: Use the filter weight vector at the current moment to estimate the noise components of the reference input vector to obtain the estimated background noise components; S33: Differentiate the internal mixed signal and the estimated background noise component in the dual-channel synchronous audio dataset to obtain a clean airflow friction sound signal.

4. The intelligent monitoring method for sand and gravel concrete refrigeration equipment according to claim 1, characterized in that, Step S4 includes: S41: Window the pure airflow friction sound signal to obtain a windowed frame sequence matrix; S42: Perform Fast Fourier Transform and Power Spectral Density Estimation on the windowed frame sequence matrix to obtain the full-band power spectral matrix; S43: Lock and extract the high-frequency aerodynamic feature range of the full-band power spectrum matrix to obtain the high-frequency acoustic signature feature spectrum.

5. The intelligent monitoring method for sand and gravel concrete refrigeration equipment according to claim 1, characterized in that, Step S5 includes: S51: Normalize the probability density of the high-frequency acoustic signature feature spectrum to obtain a normalized spectral distribution matrix. S52: Estimate the frequency disorder of the normalized spectral distribution matrix to obtain the spectral entropy eigenvector; S53: Calculate the average power intensity of the high-frequency acoustic signature feature spectrum within the feature frequency band to obtain the in-band energy feature vector; S54: The acoustic roughness index is obtained by weighted fusion and moving average filtering of the spectral entropy eigenvector and the in-band energy eigenvector.

6. The intelligent monitoring method for sand and gravel concrete refrigeration equipment according to claim 5, characterized in that, Step S6 includes: S61: Calculate the relative rate of change of the acoustic roughness index with respect to the system clean baseline fingerprint to obtain the feature offset; S62: Perform multi-phase fuzzy membership evaluation on the feature offset to obtain the fuzzy membership vector; S63: Perform maximum membership decision and time-domain de-jittering latch on the fuzzy membership vector to obtain the frosting phase state.

7. The intelligent monitoring method for sand and gravel concrete refrigeration equipment according to claim 1, characterized in that, Step S7 includes: S71: Perform logical branch mapping on the values ​​of the frosting phase state to determine the corresponding strategy mode identifier; S72: Obtain the executor target parameter vector based on the strategy mode identifier; S73: Map the physical operating parameters in the actuator target parameter vector to hardware register values ​​and append cyclic redundancy check codes to obtain optimized calibration control instructions.

8. An intelligent monitoring system for sand and gravel concrete refrigeration equipment, characterized in that, include: The signal acquisition module is used to acquire internal mixed signals, including airflow noise and background transmitted noise, and external reference signals containing only environmental background noise, through monitoring channel sensors deployed in parallel in the air duct of the chiller unit and reference channel sensors deployed outside the unit chassis. The signal synchronization module is used to perform time-stamp synchronization alignment and pre-emphasis processing on the internal mixed signal and the external reference signal to obtain a dual-channel synchronized audio dataset. The background noise cancellation module is used to adaptively cancel background noise in a dual-channel synchronous audio dataset to obtain a clean airflow friction sound signal. The high-frequency acoustic signature feature extraction module is used to perform short-time Fourier transform and frequency domain mapping on the pure airflow friction sound signal to obtain a high-frequency acoustic signature feature spectrum. The acoustic roughness analysis module is used to quantify and extract the acoustic fingerprint of micro-roughness from the high-frequency acoustic pattern feature spectrum to obtain the acoustic roughness index. The frosting state discrimination module is used to intelligently discriminate the frosting state of the acoustic roughness index based on the acoustic fingerprint evolution to determine the current frosting phase state of the evaporator based on the preset system cleaning benchmark fingerprint. The control command generation and distribution module is used to generate optimized calibration control commands based on the frosting phase state. The optimized calibration control commands are then distributed to the refrigeration unit controller to perform proactive defensive adjustments.

Images