A Health Prediction Method for HVAC Equipment Based on Audio Recognition

Abstract

This invention discloses a method for predicting the health status of HVAC equipment based on audio recognition, specifically relating to the field of HVAC equipment health status prediction technology. The method includes acquiring the rotation word and edge sound segments of a building edge computing node during a rotation switch; capturing the load sound of the switching-out equipment at the switching-out time and the load sound of the switching-in equipment at the switching-in time to generate sound segment pairs; performing a redistribution spectrum algorithm on the sound segment pairs, sampling window by window according to the sampling rate corresponding to the number of points per second, obtaining the phase difference between adjacent windows and the phase difference between adjacent frequencies through Fourier transform, writing the original time-frequency grid energy into the redistribution coordinates, and generating a switching-out spectrum and a switching-in spectrum; binding the equipment operation words before and after the rotation switch with the edge sound segments through the building edge computing node; extracting the spectral envelopes on both sides of the switching-out and switching-in sides using the redistribution spectrum algorithm and Bayesian multi-cone window spectrum estimation; determining whether abnormal sound patterns migrate with the equipment position based on the rotation migration word; and then generating a health status prediction value and maintenance priority.

Description

Technical Field

[0001] This invention relates to the field of HVAC equipment health prediction technology, and more specifically, to a method for predicting the health of HVAC equipment based on audio recognition. Background Technology

[0002] Existing HVAC equipment health prediction based on audio recognition is mostly used to detect bearing wear, impeller loosening, valve jamming and abnormal vibration in equipment such as fans, water pumps, compressors and valves in advance. In engineering, microphones or acoustic acquisition nodes are usually placed near the equipment. Edge computing nodes extract spectral energy, impact components, friction components and low-frequency vibration components, and then compare them with historical benchmarks or abnormal soundprint samples to form a health result. In large building air conditioning rooms, multiple pumps, fans or units of the same model share pipelines, supports and air ducts, and are put into operation alternately according to the main and standby rotation or load rotation method. At the same time, it is not allowed to shut down the machine separately for diagnosis on site, and it is not allowed to add any perceptible test actions. Audio data also needs to be used to complete the main judgment on the edge side. Under such operating conditions, existing audio recognition results are prone to voiceprint attribution bias. Specifically, the same sharp sound, friction sound, or low-frequency vibration sound may sometimes shift with the equipment number switch, but may sometimes appear fixed in the same pipeline section, bracket connection, or air duct reflection area. Single-point acquisition, multi-point noise reduction, and anomaly classification can only indicate that the acquired sound is abnormal, but cannot prove that the source of the anomaly belongs to the target equipment itself. In actual maintenance, there may be situations where the abnormal noise remains in the original position after the equipment is replaced according to the health result, or the abnormal noise shifts with the operating equipment after the main and backup equipment is switched, making it difficult for the prediction results to accurately point to the maintenance object. The technical problem to be solved by this application is: under edge computing conditions, how to use the main and backup rotation or load rotation process of HVAC equipment to determine whether abnormal sound patterns migrate with the equipment number, so as to distinguish between the degradation of the target equipment body and the fixed-position acoustic interference formed by pipelines, supports and air ducts. Summary of the Invention

[0003] To overcome the aforementioned deficiencies in the prior art, embodiments of the present invention provide a method for predicting the health status of HVAC equipment based on audio recognition. This method binds the equipment operation words before and after rotation switching to edge sound segments through building edge computing nodes, extracts the spectral envelopes on both sides of the switch-out and switch-in sides using a redistribution spectrogram algorithm and Bayesian multi-cone window spectral estimation, and determines whether abnormal sound patterns migrate with the equipment position based on the rotation migration word. Then, it generates a health status prediction value and maintenance priority to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the health status of HVAC equipment based on audio recognition, comprising: S1. Obtain the rotation word and edge sound segment of the building edge computing node in a rotation switch, extract the load sound of the switching-out equipment at the switching-out time, and extract the load sound of the switching-in equipment at the switching-in time to generate sound segment pairs; S2. Perform a redistribution spectrum algorithm on the sound segment pair, sample window by window according to the number of points per second corresponding to the sampling rate, obtain the phase difference between adjacent windows and the phase difference between adjacent frequencies through Fourier transform, write the original time-frequency grid energy into the redistribution coordinates, and generate the replacement spectrum and the replacement spectrum. S3. Perform Bayesian multi-cone window spectral estimation on the replaced and replaced spectra. Generate orthogonal cone windows according to the discrete long ellipsoid sequence. Generate posterior weights based on the ratio of the cone window spectral value to all cone window spectral values ​​of the same redistributed coordinate. Combine the envelopes of the replaced and replaced spectra. S4. Using the rotation word as an index on the edge side, write the envelope of the replaced spectral line into the replaced device bit and the envelope of the replaced spectral line into the replaced device bit. Take the first bit in ascending order of the sum of frequency difference, energy difference and attenuation order difference to generate the rotation migration word. S5. Generate health prediction results based on the rotation migration word. When the first result changes from the swapped-out equipment position to the swapped-in equipment position, write equipment degradation evidence. When the first result stays at the same acquisition position, write position interference record. Then, take the equipment degradation evidence as the health decline item, and take maintenance time occupation and rotation load occupation as linear constraints. Output the health prediction value and maintenance order through the simplex method.

[0005] In a preferred embodiment, S1 includes: S1-1. Obtain the device operation word formed by the building edge computing node according to the sampling order, perform differential calculation on the adjacent sampling bits by subtracting the previous bit role value from the next bit role value, write the sampling bit that changes from the load value to the exit value as the swap-out time, and write the sampling bit that changes from the standby value to the load value as the swap-in time. The role value refers to the running identity value written by the building edge computing node for a single HVAC device at each sampling point, which is used to indicate the status of a single HVAC device at that sampling point: bearing the load, exiting the load, or waiting to take over the load. The load capacity value refers to the value in the role value that indicates that the corresponding HVAC equipment has been connected to the rotating load and participated in cooling, heating or ventilation operation at this sampling point; The exit value refers to the value in the role value that indicates that the corresponding HVAC equipment has been removed from the rotating load at this sampling point and is no longer operating as the current load-bearing object; The standby value refers to the value in the role value that indicates that the corresponding HVAC equipment is not currently undertaking the rotating load at this sampling point, but is in a state where it can take over the operation; S1-2. In the equipment operation word, the swapping out equipment number is read back at the swapping out time and a swapping out side word is formed. The swapping in equipment number is read back at the swapping in time and a swapping in side word is formed. Then, the rotation word is synthesized according to the sampling order. S1-3. Based on the rotation word, read the edge sound segments, extract the sound of the switching out equipment maintaining the load value from the time of switching out to form the load sound of the switching out equipment, and extract the sound of the switching in equipment maintaining the load value from the time of switching in to form the load sound of the switching in equipment, and generate sound segment pairs by binding them according to the same rotation word. The edge sound segment refers to the sound data segment saved by the building edge computing node on the HVAC equipment site in the sampling order. The sound data segment and the equipment operation word use the same sampling bit index, so that the switching-out time and switching-in time can directly locate the corresponding sound position.

[0006] In a preferred embodiment, S2 includes: S2-1. Using the audio segment pair as input, read the sound window by window starting from the first sampling bit according to the number of points per second corresponding to the sampling rate. Move the last bit of the current window by one second to form the next window, and write the switching out load sound and the switching in load sound as the switching out window chain and the switching in window chain respectively. The sampling rate refers to the number of sampling bits written per second to the edge sound segment when the building edge computing node collects edge sound segments at equal intervals. It is used to convert one second of sound into one second of points and to read the load sound of the switching-out equipment and the load sound of the switching-in equipment window by window according to the same sampling bit scale. S2-2. Perform Fourier transform on each window in the swap-out window chain and the swap-in window chain, write the square of the modulus of the complex spectrum value as the original time-frequency grid energy, write the argument of the complex spectrum value as the phase value, and generate the swap-out time-frequency grid table and the swap-in time-frequency grid table according to the window order and frequency order. The complex spectral value refers to the frequency domain value containing real and imaginary parts formed at a frequency level after the Fourier transform of a single-window sound in the substituted window chain or the substituted window chain. The square of the modulus of the complex spectral value is used to represent the original time-frequency grid energy, and the argument of the complex spectral value is used to represent the phase value.

[0007] In a preferred embodiment, S2 further includes: S2-3. In the swapped-out and swapped-in time-frequency grid tables, the phase value of the next window at the same frequency position is subtracted from the phase value of the previous window to generate a time phase difference, and the phase value of the next frequency position within the same window is subtracted from the phase value of the previous frequency position to generate a frequency phase difference. The window order is rewritten by the time phase difference and the frequency position order is rewritten by the frequency phase difference to generate candidate redistribution coordinates. The process of rewriting the window order from the time phase difference and the frequency order from the frequency phase difference is as follows: Based on the original window order and frequency order of the time-frequency grid, the time phase difference is converted into window offset bits according to the complete phase period, and the original window order is added to the window offset bits to obtain the redistributed window order; the frequency phase difference is converted into frequency offset bits according to the complete phase period, and the original frequency order is added to the frequency offset bits to obtain the redistributed frequency order; the redistributed window order and the redistributed frequency order form the candidate redistribution coordinates. S2-4. Using the candidate redistribution coordinates as the writing positions, the original time-frequency grid energy is moved into the candidate redistribution coordinates grid by grid. The energy within the same candidate redistribution coordinate is accumulated, and after all the original time-frequency grid energy has been moved in, the replacement spectrum and the replacement spectrum are generated. The process of generating the replacement spectrum and replacement spectrum after all original time-frequency grid energies have been transferred in is as follows: taking the replacement time-frequency grid table and the replacement time-frequency grid table as objects, blank spectrum grids are established according to the candidate redistribution coordinates, and each original time-frequency grid energy is written into the corresponding blank spectrum grid; when the same candidate redistribution coordinate receives multiple original time-frequency grid energies, the accumulation is performed, and when the same candidate redistribution coordinate does not receive any original time-frequency grid energies, it is recorded as zero. Finally, the replacement spectrum and replacement spectrum are obtained by arranging them according to the window order and frequency order.

[0008] In a preferred embodiment, S3 includes: S3-1. Using the replaced spectrum and the replaced spectrum as input, read the frequency energy column in the corresponding window according to the same redistribution coordinates, and generate a discrete long ellipsoid sequence according to the length of the frequency energy column. Then, multiply each discrete long ellipsoid sequence with the frequency energy column bit by bit to form an orthogonal cone window spectrum. The frequency energy column refers to the ordered numerical column formed by reading the energy values ​​of each frequency position in the window in the frequency position order, based on the window order corresponding to the same redistribution coordinate in the outgoing or incoming spectrum, and is used to represent the distribution of sound energy along the frequency direction in the window. The process of generating a discrete long ellipsoid sequence based on the length of the frequency energy column is as follows: using the length of the frequency energy column as the matrix order, construct the tridiagonal feature matrix corresponding to the discrete long ellipsoid sequence, perform eigenvalue decomposition on the tridiagonal feature matrix, read the eigenvectors in descending order of eigenvalues, and normalize each eigenvector by the sum of squares of its elements to obtain a discrete long ellipsoid sequence. S3-2. Perform power spectrum estimation for each orthogonal cone window spectrum column, sum the cone window spectrum values ​​under the same redistribution coordinates to the coordinate spectrum sum, and generate initial posterior weights based on the proportion of a single cone window spectrum value to the coordinate spectrum sum. The cone window spectral value refers to the sum of the squared values ​​of each frequency bit in the orthogonal cone window spectral column, which is used to represent the spectral energy contribution of the corresponding discrete long ellipsoidal sequence to the frequency bit energy column in the same redistributed coordinate.

[0009] In a preferred embodiment, S3 further includes: S3-3. Calculate the envelope cost of the replaced spectrum and the replaced spectrum under the same redistribution coordinates based on the initial posterior weights. The envelope cost is formed by accumulating the absolute value of the difference in the cone window spectral values, the absolute value of the difference in the posterior weights, and the absolute value of the difference in the order of adjacent frequency positions. The first position is taken in ascending order of the envelope cost to generate the receiving cone window. S3-4. Write back the posterior weights corresponding to the receiving cone window to the same redistribution coordinates. Combine the replaced and replaced spectra according to the written-back posterior weights to synthesize the spectral envelopes, and output the replaced and replaced spectral envelopes. The process of weighted synthesis of spectral envelopes is as follows: taking the same redistribution coordinates as the processing object, the posterior weights and frequency energy columns corresponding to the receiving cone window are read, and each frequency energy value in the frequency energy column is multiplied by the posterior weight to obtain the weighted frequency energy value. The weighted frequency energy values ​​are then connected in frequency order. The weighted frequency energy values ​​under each coordinate are then connected in redistribution coordinate order to form the replaced spectral envelope and the replaced spectral envelope respectively.

[0010] In a preferred embodiment, S4 includes: S4-1. Using the rotation word as input, read the outgoing spectral envelope and the incoming spectral envelope corresponding to the same rotation word on the edge side, write the outgoing spectral envelope into the outgoing device bit, write the incoming spectral envelope into the incoming device bit, and generate the device bit envelope table. S4-2. Based on the device bit envelope table, use a single frequency bit in the swapped-out device bit as the search bit, read the same-sequence frequency bit in the swapped-in device bit, calculate the absolute value of the frequency bit difference and the absolute value of the energy difference between the two, and write the frequency-energy difference word.

[0011] In a preferred embodiment, S4 further includes: S4-3. Based on the device bit envelope table, generate the replacement attenuation sequence and the replacement attenuation sequence according to the order of energy from rising to falling, and calculate the absolute value of the order difference between the replacement attenuation sequence and the replacement attenuation sequence bit by bit, and write the attenuation difference word. S4-4. Using the frequency difference word and attenuation difference word as input, accumulate the absolute values ​​of the frequency difference, energy difference, and order difference at the same frequency position, and take the first digit of the accumulation result in ascending order to generate the round-shift word.

[0012] In a preferred embodiment, S5 includes: S5-1. Using the rotation migration word as input, read the source device bit, target device bit, and acquisition position corresponding to the first and second results. When the source device bit is the swapped-out device bit and the target device bit is the swapped-in device bit, write device degradation evidence. When the acquisition position has not changed, write position interference record and output evidence diversion table. S5-2. Based on the evidence diversion table, read the equipment degradation evidence, write the sum of the frequency difference, energy difference and attenuation order difference in the first result as the migration residual, and add the value obtained by dividing one by one and the migration residual according to the target equipment to generate a health decline item.

[0013] In a preferred embodiment, S5 further includes: S5-3. Using the health decline term as the target row coefficient and the maintenance time and shift load as the constraint row coefficients, write the maintenance selection variable, right-hand term and relaxation variable according to the target equipment to generate a simplex table. Maintenance man-hours refer to the time required for maintenance personnel to perform maintenance on the target equipment. The constraint row coefficients are written into the simplex table according to the maintenance item duration of the target equipment in the maintenance plan, and are used to limit the total maintenance man-hours that can be arranged within the same planning cycle. Rotational load occupancy refers to the load share that needs to be taken over by other HVAC equipment after the target equipment is removed from rotation due to maintenance. It is written into the constraint row coefficient of the simplex table according to the rotational load ratio corresponding to the period when the target equipment is removed, and is used to limit the occupation of the HVAC operation load-bearing capacity by the maintenance schedule. S5-4. Perform the simplex method on the simplex tableau, take the positive coefficients of the target row into the basis column, and take the first quotient of the right-hand term divided by the positive coefficients of the basis column as the out-of-base row in ascending order. Iterate until there are no positive coefficients in the target row, and output the health prediction value and maintenance order.

[0014] The technical effects and advantages of this invention are as follows: 1. This solution uses rotating migration characters to determine whether the voiceprint changes with the device location, which can distinguish between device degradation and fixed-location acoustic interference, and relatively improve the attribution of health prediction objects; 2. Bind the device operation word and the edge sound segment with the same sampling bit index so that the time of switching out and the time of switching in correspond to the sound position, thereby reducing the risk of mismatch between sound segments across devices; 3. The original time-frequency grid energy is redistributed and replaced with outgoing and incoming spectra, making the time-frequency positions of short anomalous sounds more concentrated, providing a basis for spectral envelope comparison; 4. Generating posterior weights and synthesizing spectral envelopes in Bayesian multi-cone window spectral estimation can relatively suppress fluctuations in single-window spectral estimation and improve the reliability of voiceprint acceptance judgment. 5. The frequency difference, energy difference and attenuation order difference are combined to generate the round-valued migration word, so that the voiceprint migration judgment is expanded from a single similarity to a multi-dimensional consistency judgment. Attached Figure Description

[0015] Figure 1 This is a roadmap for predicting the health status of HVAC equipment according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Refer to the instruction manual appendix Figure 1 The present invention provides a method for predicting the health status of HVAC equipment based on audio recognition, comprising: S1. Obtain the rotation word and edge sound segment of the building edge computing node in a rotation switch, extract the load sound of the switching-out equipment at the switching-out time, and extract the load sound of the switching-in equipment at the switching-in time to generate sound segment pairs; Before entering acoustic calculations, a rotation switch needs to form a data chain within the building edge computing node that can simultaneously point to the running status and sound location. This implementation uses the change in role value in the device operation word as the basis for rotation value identification, and binds the switching-out device, switching-in device, and edge sound segments along the sampling bit index, so that the sounds on both sides read by the subsequent redistribution spectrum algorithm all come from the same load handover; the implementation process includes the following steps: When the building edge computing node reads the device operation word in the sampling order, it performs the transposition boundary recognition of S1-1. The device operation word is indexed by the device number and the sampling bit. Each sampling bit is written with a role value. The role value is given by the operation configuration table. The carrying value, exit value and standby value are written as different integers respectively. For adjacent sampling bits under the same device number, the building edge computing node performs differential calculation by subtracting the previous role value from the subsequent role value. When the load value corresponding to the differential result becomes the exit value, the subsequent sampling bit is written as the swap-out time. When the standby value corresponding to the differential result becomes the load value, the subsequent sampling bit is written as the swap-in time. When multiple device numbers change their role values ​​at the same sampling position, the candidate rotation record is written in ascending order of device number. Then, the record containing both a swap-out time and a swap-in time is read from the candidate rotation record as the valid rotation switching record. If no swap-out or swap-in time is formed, no audio segment pair is generated, and the corresponding sampling bit is written into the missing cycle value record; After the effective rotation switch record is determined, the building edge computing node generates a rotation word around S1-2: first, read back the equipment number corresponding to the switch-out time in the equipment operation word, and write the equipment number, switch-out time and exit value into the switch-out side word; then read back the equipment number corresponding to the switch-in time, and write the equipment number, switch-in time and load value into the switch-in side word; then bind the switch-out side word and the switch-in side word into a rotation switch index according to the sampling order. When there is a sampling bit interval between the swap-out time and the swap-in time, the time with the earlier sampling bit is written into the first bit of the round value word, and the time with the later sampling bit is written into the last bit of the round value word; When there are multiple candidate replacement device numbers in the candidate rotation record, only the device number whose role value changes from standby value to carrying value is read, and the device number whose role value changes from standby value to exit value is discarded, so that the rotation word corresponds to a clear load handover; After the rotation word is formed, the edge sound segment binding process begins. The building edge computing node reads the edge sound segment according to S1-3 using the rotation word as the index. The edge sound segment saves the sound data according to the same sampling bit index as the device operation word. When the edge sound segment sampling bit falls between two device operation word sampling bits, it is assigned to the device operation word sampling bit with the earlier time distance; if the distances are the same, it is assigned to the later sampling bit. Read the edge sound segment backward from the moment of replacement until the next sample bit after the first time the role value of the replacement device is not the load value, and obtain the load sound of the replacement device. Read the edge sound segments from the moment of replacement until the previous sample bit before the first time the role value of the replacement device is not the load value, and obtain the load sound of the replacement device; When the number of sampling bits on both sides of the sound is inconsistent, the sound is trimmed according to the side with fewer sampling bits, and then the sound segments are generated by binding them according to the same round value word. After the above processing, the building edge computing node writes a round-shift as a calculation record containing the outgoing side character, the incoming side character, and the voice segment pair. The subsequent redistribution spectrum algorithm can use the same round-shift character to read the load sound of the outgoing equipment and the load sound of the incoming equipment, and complete the homologous voiceprint migration analysis. In practical applications: When two chilled water pumps are running in a primary / standby rotation, the building edge computing node records the change of the role value of pump 1 from the load value to the exit value at sampling position 1200, and records the change of the role value of pump 2 from the standby value to the load value at sampling position 1210. Then, the exit side word is generated with pump 1 and the entry side word is generated with pump 2. The sound of pump 1 still bearing the load before exiting and the sound of pump 2 bearing the load after entering are extracted from the edge sound segment to form a sound segment pair for subsequent acoustic calculation.

[0018] S2. Perform a redistribution spectrum algorithm on the sound segment pair, sample window by window according to the number of points per second corresponding to the sampling rate, obtain the phase difference between adjacent windows and the phase difference between adjacent frequencies through Fourier transform, write the original time-frequency grid energy into the redistribution coordinates, and generate the replacement spectrum and the replacement spectrum. Before the sound segments are entered into the redistribution spectrum calculation, the sounds on both sides need to be converted into time-frequency grids with the same window length and the same frequency scale, and then the energy dispersed in the original time-frequency grids is transferred to the actual sound source location by using phase changes. This implementation method determines the number of points per second using the sampling rate, obtains the original time-frequency grid energy and phase values ​​using Fourier transform, and then generates candidate redistribution coordinates based on the time phase difference and frequency phase difference, ultimately obtaining the replacement-out spectrum and replacement-in spectrum; the implementation process includes the following steps: When establishing a window chain along the sampling position, S2-1 takes the sound segment pair as input, the building edge computing node reads the sampling rate, the sampling rate comes from the number of sampling bits written per second in the edge sound segment acquisition configuration, and writes the number of sampling bits per second corresponding to the sampling rate as the number of points per second; Starting from the first sample bit of the load sound of the swapped-out device, read the number of points per second to form the current window, then move the last bit of the current window to the next second to form the next window, until the number of remaining sample bits is less than the number of points per second, and stop writing windows to obtain the swapped-out window chain; Perform the same window-by-window reading process on the load sound of the switching-in device to obtain the switching-in window chain; the sound with less than one second of points at the end is not written into the switching-out window chain or the switching-in window chain, so that each subsequent window has the same number of sampling bits. In the frequency domain conversion stage, S2-2 takes over the output window chain and the input window chain. The building edge computing node performs a Fourier transform on each window in the output window chain. At each frequency position, a complex spectrum value containing real and imaginary parts is obtained. The square of the modulus of the complex spectrum value is written as the original time-frequency grid energy. The argument of the complex spectrum value is written as the phase value. The time-frequency grid table is generated using window order and frequency position order as row and column indices. Similarly, perform a Fourier transform on each window in the swap-in window chain, write the original time-frequency grid energy and phase values ​​in the same frequency order, and generate the swap-in time-frequency grid table. If all the sound values ​​in a single window are zero, the original time-frequency grid energy of the corresponding frequency position is written as zero, the phase value follows the phase value of the previous window at the same frequency position, and the phase value is written as zero when there is no previous window in the first window. The candidate redistribution coordinates are calculated from the phase difference in S2-3. The building edge computing nodes read the time-frequency grid table of the swap-out and the time-frequency grid table of the swap-in, respectively. For the phase value of the next window at the same frequency position minus the phase value of the previous window, first perform phase expansion according to the complete phase period so that the difference falls between the negative half period and the positive half period. Then convert the expanded difference into the number of window offset bits and add the number of window offset bits to the original window order to obtain the redistribution window order. Subtract the previous frequency phase value from the next frequency phase value within the same window, and obtain the frequency phase difference using the same phase expansion method. Then, convert the frequency phase difference into the number of frequency offset bits, and add the number of frequency offset bits to the original frequency order to obtain the redistributed frequency order. The window offset bits and frequency offset bits are taken as the nearest integers. If the two integers are the same distance apart, the integer with the smaller absolute value is taken. When the redistribution window order crosses the window chain boundary, the boundary window order is written; when the redistribution frequency bit order crosses the frequency bit boundary, the boundary frequency bit order is written. The candidate redistribution coordinates are composed of the redistribution window order and the redistribution frequency bit order. Energy migration is completed in S2-4. The building edge computing node takes the swapped time-frequency grid as the object, establishes a blank spectrum grid according to the candidate redistribution coordinates, and writes the energy of each original time-frequency grid into the corresponding blank spectrum grid. When the same candidate redistribution coordinate receives multiple original time-frequency grid energies, the accumulation is performed. When the same candidate redistribution coordinate does not receive any original time-frequency grid energies, it is recorded as zero. Then, all blank spectrum grids are arranged in window order and frequency order to generate the replaced spectrum. The incoming time-frequency grid table is processed according to the same migration rules to generate the incoming spectrum; if the outgoing or incoming time-frequency grid table is empty, the building edge computing node stops generating the corresponding spectrum and writes the source of the empty table into the missing sound segment record to prevent the empty spectrum from entering the Bayesian multi-cone window spectrum estimation. After the above processing, the load sounds of the switching out equipment and the load sounds of the switching in equipment are uniformly converted into switching out and switching in spectra with the same window order, frequency order and energy writing rules. The energy positions of short impact sound, friction sound and resonance sound in the time-frequency plane are re-collected, and subsequent Bayesian multi-cone window spectrum estimation can read comparable frequency energy columns under the same redistribution coordinates. In practical applications: After the chilled water pumps switch shifts, the building edge computing node saves the edge sound segments at a sampling rate of 48 kHz, which means there are 48,000 sampling bits per second. The load sounds of the outgoing and incoming equipment are read window by window at 48,000 sampling bits. After Fourier transform and phase difference conversion, the original time-frequency grid energy is transferred into the candidate redistribution coordinates to form the outgoing and incoming spectra for subsequent spectral envelope calculations.

[0019] S3. Perform Bayesian multi-cone window spectral estimation on the replaced and replaced spectra. Generate orthogonal cone windows according to the discrete long ellipsoid sequence. Generate posterior weights based on the ratio of the cone window spectral value to all cone window spectral values ​​of the same redistributed coordinate. Combine the envelopes of the replaced and replaced spectra. After the replaced and replaced spectra are entered into the spectral envelope calculation, comparable frequency-energy structures need to be obtained under the same redistribution coordinates, and the uncertainty of the single-window spectrum estimation is transformed into posterior weights through Bayesian multi-cone window spectrum estimation. This implementation generates a discrete long ellipsoidal sequence based on a frequency-energy sequence, then forms a receiving cone window using an orthogonal cone window spectral sequence, cone window spectral values, and initial posterior weights, and finally synthesizes the envelopes of the replaced and replaced spectral lines; the implementation process includes the following steps: The frequency structure under the same redistribution coordinates is established starting from S3-1, and the building edge computing nodes take the swapped-out spectrum and swapped-in spectrum as inputs; Read the window sequence corresponding to the current redistribution coordinates according to the redistribution coordinate order, and read the energy value of each frequency bit in the frequency bit order within the window sequence to form the energy column of the swapped-out frequency bit and the energy column of the swapped-in frequency bit respectively. The length of the frequency energy column is used as the matrix order. The row and column indices of the tridiagonal feature matrix correspond to the frequency order. The main diagonal terms and adjacent secondary diagonal terms of the tridiagonal feature matrix are read from the tridiagonal coefficient table corresponding to the length of the frequency energy column within the edge computing node. Perform eigenvalue decomposition on the tridiagonal eigenma matrix, read all eigenvectors in descending order of eigenvalues, and normalize each eigenvector by the sum of squares of its elements to form a discrete long ellipsoid sequence. Each discrete long ellipsoid sequence is multiplied bit by bit with the energy column of the frequency position to form the orthogonal conical window spectrum of the frequency position, and multiplied bit by bit with the energy column of the frequency position to form the orthogonal conical window spectrum of the frequency position; When the frequency energy column is empty, the current redistribution coordinates are written to an empty coordinate record and do not participate in the subsequent envelope cost calculation. In S3-2, the spectral energy contribution is transformed into the initial posterior weight. The building edge computing node performs power spectrum estimation on each outgoing orthogonal cone window spectrum and each incoming orthogonal cone window spectrum, which is to square the values ​​of each frequency bit in the orthogonal cone window spectrum and then sum them up to obtain the outgoing cone window spectrum value and the incoming cone window spectrum value. The sum of all outgoing cone window spectral values ​​under the same redistribution coordinates is the sum of the outgoing coordinate spectral values, and the sum of all incoming cone window spectral values ​​is the sum of the incoming coordinate spectral values. The initial posterior weight of the replacement is obtained by dividing the value of a single replacement cone window spectrum by the sum of the replacement coordinate spectra and the initial posterior weight of the replacement is obtained by dividing the value of a single replacement cone window spectrum by the sum of the replacement coordinate spectra. When the coordinate spectrum sum is zero, the initial posterior weights of each discrete long ellipsoid sequence are written with the same value, so that the redistributed coordinates retain computable weights and avoid division by zero. The selection of the receiving cone window is completed through S3-3. The building edge computing node takes the out-of-cone window spectrum value, the incoming cone window spectrum value, the out-of-cone initial posterior weight and the incoming initial posterior weight under the same redistribution coordinate as input, and calculates the absolute value of the difference of the cone window spectrum value and the absolute value of the difference of the posterior weight for each discrete long ellipsoid sequence. The absolute value of the order difference between adjacent frequency bits is generated by the frequency bit energy column. Specifically, the energy value of the next energy bit in the adjacent frequency bit energy column of the swapped-out frequency bit is subtracted from the previous energy value. Positive values ​​are written into rising codes, negative values ​​are written into falling codes, and zero values ​​are written into level codes. The same encoding is performed on the swapped-in frequency bit energy column, and then the absolute value of the code difference between the two sides is calculated bit by bit. The absolute values ​​of the difference in spectral values ​​of the cone window, the absolute values ​​of the difference in posterior weights, and the absolute values ​​of the difference in the order of adjacent frequencies are accumulated to form the envelope cost. The first value of the envelope cost is then taken in ascending order to generate the receiving cone window. When the envelope costs are the same, the first one is selected in ascending order based on the absolute value of the posterior weight difference; if they are still the same, the first one is selected in descending order based on the eigenvalues ​​of the discrete long ellipsoidal sequence. The spectral envelope is output in S3-4. The building edge computing node writes back the posterior weights corresponding to the receiving cone window to the current redistribution coordinates, and reads the energy columns of the swapped-out and swapped-in frequencies respectively. The energy value of each frequency in the outgoing frequency energy column is multiplied by the outgoing posterior weight corresponding to the receiving cone window to obtain the outgoing weighted frequency energy value. The energy value of each frequency in the incoming frequency energy column is multiplied by the incoming posterior weight corresponding to the receiving cone window to obtain the incoming weighted frequency energy value. The energy values ​​of each weighted frequency bit are connected in frequency bit order, and then connected in redistribution coordinate order, which are written as the envelope of the replaced spectral line and the envelope of the replaced spectral line respectively; If there is an empty coordinate record in the current redistribution coordinate, the weighted frequency energy value corresponding to the redistribution coordinate is written as zero, and the next redistribution coordinate is processed. After the above processing, the outgoing and incoming spectra form spectral envelopes with posterior weights under each redistribution coordinate. The cone window spectral value, posterior weight, and adjacent frequency order jointly participate in the selection of the receiving cone window, so that the subsequent round-shifting word can read the outgoing and incoming spectral envelopes with the same coordinate source and the same weight caliber. In practical applications: After the chilled water pump shifts, both the outgoing and incoming spectra contain a set of frequency energy columns under the same redistribution coordinates. The building edge computing nodes generate a discrete long ellipsoid sequence according to the length of the frequency energy column, and obtain the initial posterior weights by the proportion of the cone window spectrum value to the coordinate spectrum. When the envelope cost corresponding to a certain receiving cone window is first in ascending order, the posterior weight of the receiving cone window is written back and used to synthesize the spectral envelopes on both sides, for subsequent judgment on whether abnormal voiceprints migrate with the device bit.

[0020] S4. Using the rotation word as an index on the edge side, write the envelope of the replaced spectral line into the replaced device bit and the envelope of the replaced spectral line into the replaced device bit. Take the first bit in ascending order of the sum of frequency difference, energy difference and attenuation order difference to generate the rotation migration word. The rotation migration term is used to advance the spectral envelopes on both sides from "acoustic morphological similarity" to a verifiable result of "whether it migrates with the rotation direction"; This implementation uses a round-robin word to uniformly index the outgoing device bit, incoming device bit, and acquisition position on the edge side. First, a device bit envelope table is formed. Then, the frequency energy difference word and attenuation difference word are calculated separately. Finally, the differences at the same frequency bit are combined into a round-robin migration word. This implementation process includes the following steps: After the rotation word enters the edge side, in S4-1, the swapped-out spectral envelope and swapped-in spectral envelope bound to the same rotation word are read first; The device position is formed by the device number and the rotating role. The spectral envelope of the swapped-out line is written into the swapped-out device position, and the spectral envelope of the swapped-in line is written into the swapped-in device position. At the same time, the acquisition node position corresponding to the edge sound segment is written as the acquisition position. The device bit envelope table uses the rotation word as an index to store the swapped-out device bit, swapped-in device bit, acquisition position, and spectral envelopes on both sides, for reading the frequency energy difference word and attenuation difference word. When the envelope of the outgoing or incoming spectral line is missing, a missing envelope record is written to the edge side, and the generation of the corresponding rotation migration word for this rotation word is stopped. In S4-2, the frequency difference word is calculated from the same-order frequency bits in the device bit envelope table. The same-order frequency bits are limited to the frequency bits corresponding to the envelopes of the replaced spectral lines and the envelopes of the replaced spectral lines in the redistribution coordinate order. On the edge side, the single frequency bit in the replaced device bit is used as the retrieval bit, and the same-order frequency bits are read in the replaced device bit. The absolute value of the frequency difference is obtained by subtracting the frequency envelope sequence of the replaced spectral lines from the frequency envelope sequence of the replaced spectral lines and taking the absolute value. The absolute value of the energy difference is obtained by subtracting the energy of the frequency envelope sequence of the replaced spectral lines from the energy of the frequency envelope sequence of the replaced spectral lines and taking the absolute value. The absolute value of the frequency difference and the absolute value of the energy difference corresponding to each search bit are written into the frequency-energy difference word according to the same frequency bit. When the input device bit lacks a frequency bit of the same sequence, the frequency bit adjacent to the search bit in the input device bit is used as the substitute frequency bit, and a substitute mark is written into the frequency-energy difference word. The attenuation order is obtained by S4-3, and the energy values ​​arranged in the redistribution coordinate order within the envelopes of the replaced spectral lines and the envelopes of the replaced spectral lines are read from the edge side respectively. For adjacent energy values, subtract the previous value from the next value. If the difference is positive, write the rising code; if the difference is negative, write the falling code; if the difference is zero, use the previous non-zero code; if the first difference is zero, write the level code. When the encoding changes from rising code to falling code, the corresponding redistribution coordinates are written into the attenuation bits, and the swap-out attenuation order and swap-in attenuation order are formed according to the order in which the attenuation bits appear. On the edge side, calculate the absolute value of the order difference between the attenuation sequence of the replaced position and the attenuation sequence of the replaced position, and write the attenuation difference word according to the same frequency position. When one side is missing an attenuation position, write the absolute value of the order difference of the corresponding frequency position on the missing side as the order value of that frequency position. The rotation migration word is synthesized in S4-4. The edge side takes the frequency-energy difference word and the attenuation difference word as input, reads the absolute value of the frequency difference, the absolute value of the energy difference and the absolute value of the order difference at the same frequency, and adds up the three differences to obtain the migration difference. All frequency shift differences are sorted in ascending order. When the cumulative results are the same, the positions are selected in ascending order according to the absolute value of the frequency difference. When the absolute values ​​of the frequency difference are still the same, the positions are selected in ascending order according to the absolute value of the energy difference. When the absolute values ​​of the energy difference are still the same, the positions are selected according to the redistribution coordinate order. The edge side writes the first result along with the source device bit, target device bit, acquisition position, absolute value of frequency difference, absolute value of energy difference, absolute value of order difference, and migration difference into the round value migration word, which is then read by the health prediction result generation step. After the above processing, the envelope of the replaced spectral line and the envelope of the replaced spectral line are converted into a round-valued migration word with device position direction, acquisition position and migration difference. Subsequent steps can determine the abnormal voiceprint as it changes from the replaced device position to the replaced device position, or stays at the same acquisition position, based on the first result. In practical applications: After the two chilled water pumps complete their main and standby shifts, the edge side writes the spectral envelope of the pump before it is removed into the pump-out device position, and writes the spectral envelope of the pump after it is connected into the pump-in device position. Then, the frequency difference, energy difference and attenuation order difference are calculated along the redistribution coordinate sequence. If the first result corresponds to the equipment position direction from pump 1 to pump 2, then the rotation migration word provides input for subsequent equipment degradation evidence.

[0021] S5. Generate health prediction results based on the rotation migration word. When the first result changes from the swapped-out equipment position to the swapped-in equipment position, write equipment degradation evidence. When the first result stays at the same acquisition position, write position interference record. Then, take the equipment degradation evidence as the health decline item, and take maintenance time occupation and rotation load occupation as linear constraints. Output the health prediction value and maintenance order through the simplex method. After the rotation migration is completed, the edge side has obtained the migration direction of the abnormal voiceprint between the device position and the acquisition position. The health prediction needs to convert this migration direction into degradation evidence that can be used for maintenance scheduling. This implementation first separates the shift migration words into equipment degradation evidence or location interference records, then converts the equipment degradation evidence into health decline items, and uses the simplex method to generate health prediction values ​​and maintenance priorities within the limits of maintenance time occupancy and shift load occupancy; the implementation process includes the following steps: Evidence triage starts from S5-1. The building edge computing node takes the round-shift migration word as input and reads the source device bit, target device bit, collection location and migration difference from the first and second results. When the source device bit is equal to the swapped-out device bit and the target device bit is equal to the swapped-in device bit, it indicates that the first result has shifted from the swapped-out device bit to the swapped-in device bit along the rotation direction, and the corresponding rotation migration word is written into the device degradation evidence on the edge side. When the source device bit and the target device bit do not form a rotation direction and the acquisition position remains at the same acquisition node position, the edge side will write the corresponding rotation migration word into the position interference record; When there are empty values ​​in the source device bit, target device bit, or acquisition location, the edge side will write the round value migration word into the missing record of the migration field, but will not write it into the evidence diversion table; The health degradation item is calculated from the equipment degradation evidence in S5-2. The edge side reads the equipment degradation evidence from the evidence distribution table and reads the frequency difference, energy difference and attenuation order difference from the first and second results. The frequency difference is expressed as the frequency order difference, the energy difference as the energy proportion difference, and the attenuation order difference as the attenuation position order difference. All three types of differences are dimensionless values. On the edge side, the three types of differences are added together and written as the migration residual. The value obtained by dividing one by one and adding the migration residual is written as the credible contribution. The credible contribution is accumulated according to the target equipment to form the health decline term. When no evidence of device degradation is found, the health decline item of the target device is set to zero, and the location interference record is only retained for checking the location of the data collection node, bracket or pipeline, and is not included in the accumulation of the health decline item. The simplex table generation stage corresponds to S5-3, with the health decline term as the target row coefficient on the edge side, and maintenance time and shift load as the constraint row coefficients; Maintenance man-hours are read from the maintenance plan record and represent the personnel working time required for one maintenance of the target equipment. Rotation load is read from the rotation load record and represents the load share that needs to be taken over by other HVAC equipment after the target equipment leaves the rotation. Each target device is written with a maintenance selection variable. The available maintenance hours within the planning period are written into the right-hand side of the first constraint, and the load that can be taken over is written into the right-hand side of the second constraint. Relaxation variables are written for each constraint row to form a simplex table. Target devices with a health decline term of zero are not written with maintenance selection variables. After the solution process enters S5-4, the simplex method is performed on the simplex tableau on the edge side. The inbound basis column is selected according to the positive coefficients of the target row. The quotient is obtained by dividing the right-hand term by the positive coefficients of the inbound basis column, and the first quotient is taken as the outbound basis row in ascending order. After completing one row transformation, recalculate the target row coefficients and continue to perform the selection of in-basic columns and out-basic rows until the target row has no positive coefficients and stop iterating. When stopped, the solution value of the maintenance selection variable is read. Target equipment with variable values ​​greater than zero is written into the maintenance priority. The maintenance priority is arranged in descending order of variable values. If the variable values ​​are the same, they are arranged in descending order of health decrease. The health prediction value is obtained by subtracting the corresponding health decline item of the target equipment from the baseline health value in the equipment ledger. When the baseline health value is missing, it is written into the health baseline missing record and the health prediction value is not output. After the above processing, the voiceprint migration results in the rotating migration words are converted into equipment degradation evidence, health decline items and executable maintenance priorities, so that the acoustic recognition results enter the corresponding prediction and optimization process; In practical applications: After the chilled water pump shifts, if the first result of the shift migration word changes from the equipment position of pump 1 to the equipment position of pump 2, the edge side writes the shift migration word into the equipment degradation evidence, and adds the frequency order difference, energy ratio difference and attenuation order difference to obtain the migration residual, and then adds the credible contribution to the health decline item of pump 2. Subsequently, a simplex table is constructed by combining the maintenance man-hour occupancy in the maintenance plan record and the rotation load occupancy in the rotation load record. After solving, the predicted health value of pump No. 2 and the corresponding maintenance priority are output.

[0022] exist Figure 1 It should be noted that: This figure illustrates the complete technical route of a method for predicting the health status of HVAC equipment based on audio recognition; The process begins by obtaining the device operation word and edge sound segment from the building edge computing node. The swap-out and swap-in times are determined through role value difference, and a rotation word and sound segment pair are generated. Then, a redistribution spectrum algorithm is performed on the sound segment pair, migrating the original time-frequency grid energy to the redistribution coordinates to obtain the swap-out and swap-in spectra. Next, Bayesian multi-cone window spectral estimation is used to generate posterior weights and a receiving cone window, synthesizing the swap-out and swap-in spectral envelopes. Finally, the spectral envelopes on both sides are written to the swap-out and swap-in device bits respectively, and the frequency difference, energy difference, and attenuation order difference are calculated to generate the rotation migration word. The diamond-shaped judgment box in the figure is used to determine whether the first result changes along the equipment position from the swapped-out equipment position to the swapped-in equipment position. If the judgment is "yes", it means that the abnormal sound pattern migrates with the equipment position and enters the calculation of equipment degradation evidence and health decline item. The health prediction value and maintenance priority are output by combining the simplex method with the maintenance time occupation and the rotation load occupation. If the judgment is "no", it is written into the position interference record and does not participate in the accumulation of health decline item. In the diagram, solid arrows indicate the sequential transmission of data and calculation results, while solid arrows marked "yes" or "no" indicate the branch flow corresponding to the judgment result.

[0023] Working Principle: This scheme first uses building edge computing nodes to read the equipment operation words and edge sound segments during HVAC equipment rotation. It determines the switch-out and switch-in times through role value difference, and extracts the load sound before the switch-out equipment exits and the load sound after the switch-in equipment enters, forming sound segment pairs under the same rotation. Then, it performs a spectrum redistribution algorithm on the sound segment pairs, migrating the sound energy originally dispersed in the time-frequency grid to the actual sound source location, generating switch-out and switch-in spectra. Finally, it uses Bayesian multi-cone window spectrum estimation to extract the posterior weighted sound energy from both spectra. The spectral envelope allows for more reliable alignment and comparison of short abnormal noises, friction sounds, or resonance sounds. Next, the swapped-out spectral envelope is written to the swapped-out device position, and the swapped-in spectral envelope is written to the swapped-in device position. Frequency difference, energy difference, and attenuation order difference are calculated to generate a rotation migration word. Finally, based on the rotation migration word, it is determined whether the abnormal sound signature migrates along the device position or remains at the acquisition position. The former is written as equipment degradation evidence and used for health prediction, while the latter is written as location interference records. Combined with maintenance time occupancy and rotation load occupancy, the health prediction value and maintenance priority are output using the simplex method. In a large building's air conditioning room, when two chilled water pumps operate in a primary / standby mode, the edge computing node records the change of pump 1 from load-bearing to load-free and pump 2 from standby to load-bearing. Simultaneously, it captures the ambient sound before pump 1 is deactivated and after pump 2 is activated. If, after redistribution spectral mapping and Bayesian multi-cone window spectral estimation, the same anomalous spectral line is found to change with the equipment position from pump 1 to pump 2, the system writes this anomaly into the equipment degradation evidence and includes pump 2 in the health decline calculation. If the anomalous spectral line remains at the same sampling location, it indicates that the problem is more likely from fixed acoustic interference near the pipes, supports, or sampling point, and does not directly reduce equipment health. In this way, maintenance personnel receive not only abnormal sound alerts but also clear health predictions and maintenance priorities.

[0024] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the health status of HVAC equipment based on audio recognition, characterized in that, include: S1. Obtain the rotation word and edge sound segment of the building edge computing node in a rotation switch, extract the load sound of the switching-out equipment at the switching-out time, and extract the load sound of the switching-in equipment at the switching-in time to generate sound segment pairs; S2. Perform a redistribution spectrum algorithm on the sound segment pair, sample window by window according to the number of points per second corresponding to the sampling rate, obtain the phase difference between adjacent windows and the phase difference between adjacent frequencies through Fourier transform, write the original time-frequency grid energy into the redistribution coordinates, and generate the replacement spectrum and the replacement spectrum. S3. Perform Bayesian multi-cone window spectral estimation on the replaced and replaced spectra. Generate orthogonal cone windows according to the discrete long ellipsoid sequence. Generate posterior weights based on the ratio of the cone window spectral value to all cone window spectral values ​​of the same redistributed coordinate. Combine the envelopes of the replaced and replaced spectra. S4. Using the rotation word as an index on the edge side, write the envelope of the replaced spectral line into the replaced device bit and the envelope of the replaced spectral line into the replaced device bit. Take the first bit in ascending order of the sum of frequency difference, energy difference and attenuation order difference to generate the rotation migration word. S5. Generate health prediction results based on the rotation migration word. When the first result changes from the swapped-out equipment position to the swapped-in equipment position, write equipment degradation evidence. When the first result stays at the same acquisition position, write position interference record. Then, take the equipment degradation evidence as the health decline item, and take maintenance time occupation and rotation load occupation as linear constraints. Output the health prediction value and maintenance order through the simplex method.

2. The method for predicting the health status of HVAC equipment based on audio recognition according to claim 1, characterized in that: S1 includes: S1-1. Obtain the device operation word formed by the building edge computing node according to the sampling order, perform differential calculation on the adjacent sampling bits by subtracting the previous bit role value from the next bit role value, write the sampling bit that changes from the load value to the exit value as the swap-out time, and write the sampling bit that changes from the standby value to the load value as the swap-in time. S1-2. In the equipment operation word, the swapping out equipment number is read back at the swapping out time and a swapping out side word is formed. The swapping in equipment number is read back at the swapping in time and a swapping in side word is formed. Then, the rotation word is synthesized according to the sampling order. S1-3. Based on the rotation word, read the edge sound segments, extract the sound of the switching out equipment maintaining the load value from the time of switching out to form the load sound of the switching out equipment, and extract the sound of the switching in equipment maintaining the load value from the time of switching in to form the load sound of the switching in equipment, and generate sound segment pairs by binding them according to the same rotation word.

3. The method for predicting the health status of HVAC equipment based on audio recognition according to claim 2, characterized in that: S2 includes: S2-1. Using the audio segment pair as input, read the sound window by window starting from the first sampling bit according to the number of points per second corresponding to the sampling rate. Move the last bit of the current window by one second to form the next window, and write the switching out load sound and the switching in load sound as the switching out window chain and the switching in window chain respectively. S2-2. Perform Fourier transform on each window in the swap-out window chain and the swap-in window chain, write the square of the modulus of the complex spectrum value as the original time-frequency grid energy, write the argument of the complex spectrum value as the phase value, and generate the swap-out time-frequency grid table and the swap-in time-frequency grid table according to the window order and frequency order.

4. The method for predicting the health status of HVAC equipment based on audio recognition according to claim 3, characterized in that: S2 also includes: S2-3. In the swapped-out and swapped-in time-frequency grid tables, the phase value of the next window at the same frequency position is subtracted from the phase value of the previous window to generate a time phase difference, and the phase value of the next frequency position within the same window is subtracted from the phase value of the previous frequency position to generate a frequency phase difference. The window order is rewritten by the time phase difference and the frequency position order is rewritten by the frequency phase difference to generate candidate redistribution coordinates. S2-4. Using the candidate redistribution coordinates as the writing positions, the original time-frequency grid energy is transferred into the candidate redistribution coordinates grid by grid. The energy within the same candidate redistribution coordinate is accumulated, and after all the original time-frequency grid energy has been transferred in, the outgoing spectrum and the incoming spectrum are generated.

5. The method for predicting the health status of HVAC equipment based on audio recognition according to claim 4, characterized in that: S3 includes: S3-1. Using the replaced spectrum and the replaced spectrum as input, read the frequency energy column in the corresponding window according to the same redistribution coordinates, and generate a discrete long ellipsoid sequence according to the length of the frequency energy column. Then, multiply each discrete long ellipsoid sequence with the frequency energy column bit by bit to form an orthogonal cone window spectrum. S3-2. Perform power spectrum estimation for each orthogonal cone window spectrum, sum the cone window spectrum values ​​under the same redistribution coordinates to the coordinate spectrum sum, and generate initial posterior weights based on the proportion of a single cone window spectrum value to the coordinate spectrum sum.

6. The method for predicting the health status of HVAC equipment based on audio recognition according to claim 5, characterized in that: S3 also includes: S3-3. Calculate the envelope cost of the replaced spectrum and the replaced spectrum under the same redistribution coordinates based on the initial posterior weights. The envelope cost is formed by accumulating the absolute value of the difference in the cone window spectral values, the absolute value of the difference in the posterior weights, and the absolute value of the difference in the order of adjacent frequency positions. The first position is taken in ascending order of the envelope cost to generate the receiving cone window. S3-4. Write back the posterior weights corresponding to the receiving cone window to the same redistribution coordinates. Combine the replaced and replaced spectra into spectral envelopes by weighting them according to the written-back posterior weights. Output the replaced and replaced spectral envelopes.

7. The method for predicting the health status of HVAC equipment based on audio recognition according to claim 6, characterized in that: S4 includes: S4-1. Using the rotation word as input, read the outgoing spectral envelope and the incoming spectral envelope corresponding to the same rotation word on the edge side, write the outgoing spectral envelope into the outgoing device bit, write the incoming spectral envelope into the incoming device bit, and generate the device bit envelope table. S4-2. Based on the device bit envelope table, use a single frequency bit in the swapped-out device bit as the search bit, read the same-sequence frequency bit in the swapped-in device bit, calculate the absolute value of the frequency bit difference and the absolute value of the energy difference between the two, and write the frequency-energy difference word.

8. The method for predicting the health status of HVAC equipment based on audio recognition according to claim 7, characterized in that: S4 also includes: S4-3. Based on the device bit envelope table, generate the replacement attenuation sequence and the replacement attenuation sequence according to the order of energy from rising to falling, and calculate the absolute value of the order difference between the replacement attenuation sequence and the replacement attenuation sequence bit by bit, and write the attenuation difference word. S4-4. Using the frequency difference word and attenuation difference word as input, accumulate the absolute values ​​of the frequency difference, energy difference, and order difference at the same frequency position, and take the first digit of the accumulation result in ascending order to generate the round-shift word.

9. The method for predicting the health status of HVAC equipment based on audio recognition according to claim 8, characterized in that: S5 includes: S5-1. Using the rotation migration word as input, read the source device bit, target device bit, and acquisition position corresponding to the first and second results. When the source device bit is the swapped-out device bit and the target device bit is the swapped-in device bit, write device degradation evidence. When the acquisition position has not changed, write position interference record and output evidence diversion table. S5-2. Based on the evidence diversion table, read the equipment degradation evidence, write the sum of the frequency difference, energy difference and attenuation order difference in the first result as the migration residual, and add the value obtained by dividing one by one and the migration residual according to the target equipment to generate a health decline item.

10. A method for predicting the health status of HVAC equipment based on audio recognition according to claim 9, characterized in that: S5 also includes: S5-3. Using the health decline term as the target row coefficient and the maintenance time and shift load as the constraint row coefficients, write the maintenance selection variable, right-hand term and relaxation variable according to the target equipment to generate a simplex table. S5-4. Perform the simplex method on the simplex tableau, take the positive coefficients of the target row into the basis column, and take the first quotient of the right-hand term divided by the positive coefficients of the basis column as the out-of-base row in ascending order. Iterate until there are no positive coefficients in the target row, and output the health prediction value and maintenance order.

Images