Early warning and diagnosis method for hidden failure of fan coil motor

Abstract

The present application relates to the technical field of motor fault diagnosis, and particularly relates to a fan coil motor implicit fault early warning and diagnosis method, which is as follows: fan coil motor vibration signals are collected and running states are labeled to construct a training sample set; the optimal fault cycle is determined by using time domain repeated impact law, the same phase position is synchronously enhanced, and impact enhanced vibration samples are obtained; singular spectrum decomposition is used to separate the dominant structure, and reconstructed vibration samples and statistical feature vectors are output; a double-input gated diagnosis network is constructed, the reconstructed samples are used as the main input, and the statistical features are used as the auxiliary input, diagnosis feature vectors are obtained through fusion and gated compression, the prediction results of normal and early fault states are output through class weighting training, the trained network is deployed in an online monitoring system, and motor state real-time monitoring and early fault accurate diagnosis are realized. Through targeted time domain same phase synchronous enhancement and singular spectrum decomposition separation, accurate early warning and real-time diagnosis can be realized.

Description

Technical Field

[0001] This invention relates to the field of motor fault diagnosis technology, and in particular to a method for early warning and diagnosis of latent faults in fan coil motors. Background Technology

[0002] As terminal devices in air conditioning systems, fan coil units operate at low speeds for extended periods. Under these conditions, vibrations triggered by early faults such as bearing pitting, poor lubrication, or slight rotor imbalance exhibit extremely low energy, short duration, and unstable repetition cycles. These subtle fault characteristics are easily masked by the motor's fundamental frequency vibration, structural resonance, and broadband noise in the environment. Furthermore, the actual operating environment of fan coil unit motors is complex, with potential slight speed fluctuations. The number of normal samples in the collected vibration signals far exceeds the number of early fault samples, indicating a significant class imbalance. Therefore, a method is needed that can enhance weak periodic impacts against a strong noise background, adaptively separate fault components from background disturbances, and stably distinguish between normal and early fault states even under sample imbalance conditions.

[0003] Existing technologies objectively suffer from the following shortcomings: conventional preprocessing methods such as mean removal, standardization, or fixed bandpass filtering are difficult to simultaneously preserve weak impacts, suppress random noise, and ensure subsequent classification stability. Weak impacts generated by early faults are easily submerged by the fundamental frequency and environmental noise. Methods that determine the fault period based on a single spectral peak are easily affected by instantaneous noise and slight fluctuations in rotational speed, leading to unstable period estimation and an inability to reliably capture the impact patterns of early faults with weak repetition periods. Using fixed-band filtering for fault component separation is difficult to adapt to fault band shifts under different operating conditions, easily weakening effective components and failing to adaptively separate stable fault structures from background disturbances. Relying solely on one-dimensional temporal convolution to process the reconstructed signal does not fully utilize the overall energy level, residual fluctuation characteristics, and sample imbalance problem, making it difficult to effectively extract and distinguish the stable differences between weak early faults and normal states.

[0004] Therefore, this invention proposes an early warning and diagnosis method for latent faults in fan coil motors to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes an early warning and diagnosis method for latent faults in fan coil motors. This invention achieves accurate early warning and real-time diagnosis through targeted time-domain in-phase synchronous enhancement and singular spectrum decomposition and separation.

[0006] The technical solution of this invention is an early warning and diagnosis method for latent faults in fan coil unit motors, which operates as follows: Vibration signals from fan coil unit motors are collected and organized into training samples. The operating status of each sample is labeled to form a set of labeled training samples. The optimal fault period is determined by utilizing the repetitive impact pattern of training samples in the time domain. Synchronous reinforcement is then performed on the same phase position according to the optimal fault period to obtain impact-enhanced vibration samples. Singular spectral decomposition is used to separate the dominant structure of the impact-enhanced vibration sample, and the reconstructed vibration sample and statistical feature vector are output. The reconstructed vibration samples are used as the main input and the statistical feature vectors are used as the auxiliary input to the constructed dual-input gated diagnostic network. After fusion, the diagnostic feature vectors are obtained through gating compression. Then, the prediction results of normal state and early fault state are output through class weighted training. The trained dual-input gated diagnostic network is deployed in the online monitoring system of fan coil motor to monitor the status of fan coil motor in real time.

[0007] The training sample set contains all training samples and their corresponding labels, as follows: Using vibration sensors Vibration signals from the fan coil motor housing or bearing housing are collected at a sampling frequency. The continuously collected vibration signals are divided into multiple training samples by a fixed length. Each training sample represents a continuous vibration amplitude sequence obtained from one segmentation. Each training sample is labeled, and the label is denoted as . , Indicates the running status of the training samples. Indicates a normal state. This indicates an early stage of a fault.

[0008] The process of determining the optimal failure period is as follows: The absolute values ​​of the training samples are taken point by point to obtain an absolute value vibration sequence. The location of local maxima is detected in the absolute value vibration sequence, based on the equipment rotation speed range and sampling frequency. Set a candidate period range, select a reference impact position from the local maximum positions, and for each theoretical impact position, if an actual local maximum position is found within the allowable offset range, record the absolute value amplitude corresponding to the actual local maximum position. After traversing all theoretical impact positions, divide the cumulative amplitude by the number of theoretical impact positions to obtain the period matching score corresponding to the current candidate period. Calculate the period matching score for each candidate period within the candidate period range, and select the candidate period with the largest period matching score from all candidate periods as the optimal fault period.

[0009] The calculation process for the impact-enhanced vibration sample is as follows: The optimal fault cycle is moduloed according to the sampling point sequence number, and the training samples are divided into phase groups. Each phase group represents the relative position of the sampling point within a cycle, and the phase group number to which each sampling point belongs is its modulo of the optimal fault cycle. Then, the phase energy of each phase group is counted, and the absolute values ​​of all sampling points in the phase group are averaged to obtain the phase energy. Then, all phase energies are normalized to the range of 0 to 1 to obtain the phase enhancement weight. Then, the overall enhancement intensity is determined according to the cycle matching score corresponding to the optimal candidate cycle. The overall enhancement intensity represents the overall amplification degree of all phase enhancement weights applied to the training samples. Finally, multiplicative enhancement is performed on the training samples point by point to obtain the impact-enhanced vibration samples.

[0010] The embedding length is set according to the optimal fault cycle, the trajectory matrix is ​​constructed and singular value decomposition is performed, and then the number of principal components to be retained is automatically determined by the decay inflection point or cumulative energy threshold of the singular value sequence. Using impact-enhanced vibration samples as input, a trajectory matrix is ​​constructed. The trajectory matrix represents the trajectory of the impact-enhanced vibration samples along a length of... The two-dimensional matrix is ​​formed by expanding the sliding window point by point. Indicates the embedding length; Singular value decomposition is performed on the trajectory matrix to obtain several singular values ​​and corresponding singular components. The singular values ​​are sorted from largest to smallest. The larger the singular value, the greater the contribution of the corresponding component to the overall structure of the trajectory matrix. Then, the number of components to be retained is determined based on the magnitude of the change in singular values, and then the number of dominant singular components to be retained is determined.

[0011] The retained dominant singular components are superimposed onto the impact-enhanced vibration samples to obtain reconstructed vibration samples. The statistical feature vector is then determined based on the residual between the enhanced vibration samples and the reconstructed vibration samples. The retained dominant singular components are superimposed on the impact-enhanced vibration samples to obtain an approximate trajectory matrix. Then, the approximate trajectory matrix is ​​averaged in the opposite direction to obtain a reconstructed vibration sample. Finally, the residual vibration sample is obtained based on the difference between the impact-enhanced vibration sample and the reconstructed vibration sample. The residual vibration sample contains background noise, random disturbances, and unstable components not explained by the dominant components. Time-domain statistics are extracted from the reconstructed vibration samples and the residual vibration samples respectively. The root mean square value, peak factor, impulse factor and kurtosis are calculated for the reconstructed vibration samples and the residual vibration samples respectively. The eight statistics are concatenated in a fixed order to obtain the statistical feature vector.

[0012] The dual-input gating diagnostic network includes a timing input branch and a statistical input branch; The structure of the timing input branch and the forward propagation process are as follows: The temporal input branch has two layers of one-dimensional convolutions: the first layer has 16 kernels, and the second layer has 32 kernels. Each layer is followed by an activation function and max pooling. Finally, global average pooling outputs a 32-dimensional temporal feature vector. The reconstructed vibration sample is input into the temporal input branch. The reconstructed vibration sample contains the main fault impact structure retained after periodic enhancement and singular spectrum decomposition. It is then passed through the first convolutional layer and the second convolutional layer in sequence. Global average pooling is performed on the multi-channel feature map output by the second convolutional layer to obtain the temporal feature vector.

[0013] The statistical input branch contains a fully connected mapping layer, which can be followed by a normalization layer and a non-linear activation function. The statistical feature vector is input into the statistical input branch and mapped to a statistical feature vector.

[0014] The temporal feature vector and the statistical feature vector are concatenated by dimension to obtain the fused feature vector. The first linear transformation is performed on the fused feature vector to obtain the intermediate feature vector. After the linear transformation, a nonlinear activation function is applied to generate gating weights based on the intermediate feature vector. The intermediate feature vector is then weighted dimension by dimension to obtain the gated feature vector. The second linear transformation is performed on the gated feature vector to obtain the diagnostic feature vector.

[0015] The diagnostic feature vector is processed using Softmax to obtain the probabilities of normal and early failure. During training, class weights are calculated based on the number of samples, and weighted cross-entropy loss is used for training constraints. The specific operation is as follows: The diagnostic feature vector is input into the output layer to obtain a 2D classification logic value vector, which corresponds to the unnormalized scores of the normal state and the early fault state, respectively. Then, a Softmax transformation is performed on it to obtain the predicted probabilities of the normal state and the early fault state. The class weights are calculated based on the number of normal samples and the number of early fault samples in the training sample set. The weighted cross-entropy loss is calculated based on the class weights, and the dual-input gating diagnostic network is trained by backpropagation using this loss. After training, new sample data is input into the trained network, and the early fault state prediction probability is output. The early fault state prediction probability is compared with the preset warning threshold, and the warning result is output.

[0016] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: This invention discloses an early warning and diagnosis method for latent faults in fan coil motors. It searches for the optimal fault period most consistent with the repetitive impact pattern within a candidate period range, then groups the original vibration signal by phase according to this period, statistically analyzes the energy of each phase group and normalizes it into a phase enhancement weight. Simultaneously, it calculates the overall enhancement coefficient based on the period matching score, and finally performs multiplicative enhancement on each sampling point to synchronously amplify the periodic impact while preventing the accumulation of random noise. An upper limit for the enhancement factor is set to avoid local abnormal amplification. The optimal fault period is directly determined using the repetitive impact pattern in the time domain. Local maxima are detected in the absolute value vibration sequence, and the local maxima with the largest amplitude are used as the reference impact position. Theoretical impact positions are generated according to the candidate periods, and actual local maxima are searched within the allowable left and right offset range. The accumulated matched amplitudes are divided by the number of theoretical positions to obtain the period matching score. The candidate period with the highest score is selected as the optimal fault period, which can tolerate slight fluctuations in rotational speed. Based on the optimal fault cycle, the embedding length is adaptively set to construct the trajectory matrix and singular value decomposition is performed. The number of principal components to be retained is automatically determined by the decay inflection point or cumulative energy threshold of the singular value sequence. The dominant singular components are superimposed and averaged in anti-angle to obtain the reconstructed vibration sample. At the same time, the residual vibration sample is obtained from the difference. Eight statistics, namely root mean square value, peak factor, impulse factor and kurtosis, are extracted and concatenated into a statistical feature vector to achieve adaptive separation of fault components and background disturbances. A dual-input gated diagnostic network is constructed. The reconstructed vibration sample is input into two layers of one-dimensional convolution to extract temporal features. The statistical feature vector is mapped to auxiliary features through a fully connected layer. After concatenation, the intermediate features are obtained by the first linear transformation. Gating weights are generated in parallel and multiplied element-wise for adaptive filtering. The diagnostic features are obtained by the second linear transformation. At the same time, the class weights are calculated based on the number of normal samples and early fault samples. Weighted cross-entropy loss is used for training to improve the discrimination stability of weak faults. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0019] Figure 2 This is a vibration signal diagram of a fan coil unit motor under normal operating conditions.

[0020] Figure 3 This is a vibration signal diagram of a fan coil motor during an early bearing failure.

[0021] Figure 4 Line graph of period matching score for each candidate period.

[0022] Figure 5 Comparison images before and after periodic synchronous local impact enhancement. Detailed Implementation

[0023] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0024] Example 1 like Figure 1 As shown, a method for early warning and diagnosis of latent faults in fan coil unit motors includes the following steps: S1. Collect vibration signals of fan coil unit motors and organize them into training samples, and label the operating status of each sample to form a set of labeled training samples; S2. Determine the optimal fault period by utilizing the repetitive impact pattern of the training samples in the time domain, and synchronously enhance the same phase position according to the optimal fault period to obtain the impact-enhanced vibration sample. S3. Use singular spectral decomposition to separate the dominant structure of the impact-enhanced vibration sample and output the reconstructed vibration sample and statistical feature vector. S4. The reconstructed vibration samples are used as the main input and the statistical feature vector is used as the auxiliary input to the constructed dual-input gated diagnostic network. After fusion, the diagnostic feature vector is obtained through gate compression. Then, the prediction results of normal state and early fault state are output through class weighted training. S5. Deploy the trained dual-input gating diagnostic network in the online monitoring system for fan coil motors to monitor the status of fan coil motors in real time.

[0025] In a specific implementation, step S1, the collection and organization of vibration samples from the fan coil unit motor, is as follows: When a fan coil unit motor operates at low speed, the vibration impact caused by early faults is short in duration, weak in repetition period, and low in energy. It is easily masked by the fundamental frequency, structural vibration, and environmental noise. Conventional preprocessing methods such as mean removal, standardization, or fixed bandpass filtering often fail to simultaneously preserve weak impacts, suppress random noise, and ensure classification stability. This invention collects vibration signals from the fan coil unit motor and organizes these signals into fixed-length training samples, providing a unified input for subsequent period enhancement, fault component reconstruction, and classification training. The specific steps are as follows: S1.1 Collect vibration signals from the fan coil unit motor housing or bearing housing. An accelerometer can be used as the vibration sensor, and the sampling frequency is denoted as [missing information]. , This represents the number of vibration sampling points collected per unit time, in Hz, for example, 2000Hz. Further, the continuously acquired vibration signal is divided into multiple original vibration samples by a fixed length. Original vibration sample This represents the continuous vibration amplitude sequence obtained from a single segmentation, with its length denoted as . , This indicates the number of sampling points contained in a single sample, for example, 1024. Thus, each original vibration sample... The size is .

[0026] S1.2, For each original vibration sample Label the work, and record the label as follows: , Indicates the current original vibration sample The corresponding running states, among which, Indicates a normal state. This indicates an early stage of failure. Early stage failures can be determined from historical maintenance records, disassembly and inspection results, manual verification results, or test results under known failure conditions.

[0027] S1.3, All original vibration samples With corresponding tags The training sample set is formed, and the total number of training samples is denoted as . The subsequent steps are performed on each original vibration sample. The cycle enhancement and fault component reconstruction are performed separately, and then the processing results are fed into the classification network.

[0028] Figure 2 The graph shows the vibration signal characteristics of a fan coil unit motor under normal operating conditions. The horizontal axis represents time in seconds, and the vertical axis represents vibration amplitude in meters per square second. Figure 2 The blue curve in the middle exhibits a regular sinusoidal wave pattern with relatively uniform amplitude and no obvious sudden impacts or peaks. There are only slight random fluctuations in the background, corresponding to the fundamental vibration of the motor at a frequency of about 30 Hz and a small amount of environmental noise.

[0029] Figure 3 This refers to the vibration signal when the fan coil unit motor experiences early bearing failure. Figure 3 The horizontal axis represents time (seconds), and the vertical axis represents amplitude (meters per square second). Against the background of the overall sinusoidal fluctuation, the red curve shows several instantaneous spikes. However, the height of some spikes is similar to the noise peak, and their positions are not strictly equidistant. It is difficult to confirm them one by one with the naked eye, indicating that the impact energy generated by the early fault is low and the repetition period is weak, making it easy to be submerged by frequency conversion components and environmental noise.

[0030] In a specific implementation, step S2, the local impact enhancement processing based on periodic priors, is as follows: The weak impacts generated by early faults in fan coil unit motors are usually not completely random, but correspond to the rotational frequency, bearing fault characteristic frequency, or their related harmonics. Therefore, they exhibit an approximately repetitive impact structure in the time domain. Conventional peak amplification, envelope spectrum enhancement, or global energy enhancement methods easily amplify both sporadic noise spikes and actual fault impacts, leading to false amplification. This invention first searches for the optimal fault period within the candidate period range that best matches the repetitive impact pattern. Then, it synchronously amplifies the in-phase positions according to the optimal fault period, amplifying the periodic impacts while preventing the continuous accumulation of unstable random noise. The specific steps are as follows: S2.1 Determination of Optimal Failure Cycle Traditional methods of determining the fault period based solely on a single spectral peak are easily affected by instantaneous noise and slight fluctuations in rotational speed, leading to unstable period estimations. This invention directly utilizes the original vibration samples. The optimal fault period is determined by the repetitive impact pattern in the time domain. The specific steps are as follows: (1) For the original vibration sample By taking the absolute value point by point, the absolute value vibration sequence is obtained. Absolute value vibration sequence This is used to simultaneously cover both positive and negative impacts, avoiding missed peak values ​​due to different impact polarities. Furthermore, in absolute value vibration sequences... In the detection of local maxima, specifically, for the first... If the absolute value of the nth sampling point is greater than both the previous and the next sampling point, then the nth sampling point will be... Each sampling point is recorded as a local maximum location. To avoid mistaking minute noise fluctuations for impact peaks, an amplitude filtering condition is added to the local maxima. In one implementation, only vibration sequences with absolute values ​​greater than the absolute value are retained. The location of a local maximum at 0.5 times the standard deviation.

[0031] (2) Based on the equipment speed range and sampling frequency Define the candidate period range, and denot the optimal failure period as . , This indicates the impact of the fault in the original vibration sample. The repetition interval is expressed as the number of sampling points; if the fan coil unit motor operates between 10Hz and 50Hz, the sampling frequency is... If we take 2000Hz, the candidate period range can be 40 to 200 sampling points. This range comes from the conversion result of "sampling frequency divided by candidate frequency", and can also be corrected by combining historical fault data.

[0032] In one embodiment, for example, the historical maintenance record of a fan coil unit motor shows that the characteristic frequency of the bearing outer ring fault is approximately 25Hz, then the corresponding period is approximately With a sampling point, the candidate period range can be adjusted to 75-85, and the optimal fault period can be searched only within this small interval, which can both speed up the calculation and eliminate interference from irrelevant periods.

[0033] It should be noted that the candidate period refers to the interval length (in units of sampling points) at which the fault impact may recur in the original vibration signal. Since the rotational frequency of the fan coil unit motor or the characteristic frequency of bearing faults usually fall within a known range (e.g., 10-50Hz), based on the sampling frequency... The corresponding period range can be calculated: Period = Sampling frequency / Frequency, i.e. One sampling point, With a sampling point, the candidate period range is set to 40-200, indicating that the repetition interval of the suspected actual fault impact may fall within this range. It should also be noted that the optimal fault period... This refers to selecting the period with the highest score within the candidate period range by calculating the degree of consistency between each candidate period and the actual repetitive pattern of the impact in the signal (i.e., period matching score). This period represents the true interval of the most significant and stable repetitive impact in the signal. It should also be noted that setting a reasonable candidate period search interval avoids blindly searching in impossible ranges (such as 1 sampling point or 1000 sampling points), thereby reducing computational load and improving the reliability of period estimation. Furthermore, if the typical period of a certain fault is known from historical fault data, this range can be narrowed or corrected.

[0034] (3) Calculate the period matching score for each candidate period within the candidate period range, and record the period matching score as follows: , Indicates candidate period The degree of consistency with the repetitive impact pattern. In specific implementation, a reference impact position is first selected from the local maxima positions. The reference impact position can be the local maximum position with the largest amplitude. Then, starting from this reference impact position, according to the current candidate period... A set of theoretical impact positions is generated forward and backward along the time axis; then, with each theoretical impact position as the center, the actual local maximum position is searched within the allowable left and right offset range; the allowable offset range is used to tolerate slight fluctuations in rotational speed, for example, the current candidate period can be taken. 1 / 4 of it.

[0035] In one embodiment, for example, suppose a candidate period has been selected. There are 10 sampling points, and the reference impact location with the largest amplitude was selected from the local maxima. The sampling point number is 1. Starting from this position, according to the cycle Theoretical impact positions are generated forward and backward: Looking backward (future direction): , , ...; forward (past direction): , (Ignore samples outside the starting point); Allowed offset range is... One sampling point; for the theoretical location 230, in The system searches for the location of the actual local maximum within the interval. If there is an actual maximum within the interval (e.g., at 228), the match is considered successful, and the amplitude at that point is recorded. If there is no maximum within the interval, the match fails, and the amplitude is recorded as 0. Based on this, it can tolerate slight fluctuations in rotational speed. Actual fault impacts may not strictly occur on integer multiples of the cycle. Allowing a certain offset can more robustly evaluate cycle consistency.

[0036] (4) For each theoretical impact position, if an actual local maximum position is found within the allowable offset range, record the absolute value amplitude corresponding to that actual local maximum position; otherwise, record it as 0. After traversing all theoretical impact positions, divide the cumulative amplitude by the number of theoretical impact positions to obtain the current candidate period. Corresponding period matching score .

[0037] (5) Select the period matching score from all candidate periods. The largest candidate period is used as the optimal failure period. For example, if the period matching score corresponding to candidate period 80 is significantly higher than that of candidate periods 79, 81, and others, then the 80 sampling points are determined as the optimal fault period. .

[0038] It should be noted that the absolute value vibration sequence is used in the periodic assessment phase. Peak search is performed to capture both positive and negative impacts simultaneously, while the subsequent enhancement stage still directly targets the original vibration sample. The original amplitude values ​​are processed to preserve the original vibration samples. The phase relationship and positive and negative waveform directions in the waveform.

[0039] In one embodiment, the optimal fault period is determined, which is the most significant repetitive impact interval in the signal. The experimental configuration is as follows: within a candidate period range of 40 to 200 sampling points (corresponding to frequencies of 50 Hz to 10 Hz), for each candidate period, the theoretical impact position is generated according to the local maximum value with the largest amplitude, with reference to the local maximum value of the largest amplitude. The actual maximum value is then searched within the allowable offset range to the left and right of each theoretical position. The matched amplitudes are accumulated and divided by the number of theoretical positions to obtain the period matching score. Figure 4The horizontal axis represents the candidate period (number of sampling points), and the vertical axis represents the period matching score. The blue curve shows a significant peak near period 80, with a score significantly higher than that of adjacent periods, indicating that the actual repetitive impact interval in the signal is closest to 80 sampling points. Therefore, the optimal fault period is determined to be 80.

[0040] S2.2, Periodic Synchronization of Localized Impact Enhancement Determine the optimal failure cycle Then, if the entire original vibration sample is directly analyzed... If a uniform scale is applied, background vibrations and random spikes will also be amplified. This invention is based on the optimal fault cycle. For the original vibration sample Phase grouping is performed to give greater weight to phase positions that repeatedly show high responses across multiple periods, thereby highlighting the periodic shock structure. The specific steps are as follows: (1) Optimal fault cycle according to sampling point sequence number Take the mold and extract the original vibration sample. Divided into There are 1 phase group, and the phase group number is denoted as . , This indicates the relative position of the sampling point within one period, with a value ranging from 0 to... For the first Each sampling point belongs to a phase group numbered " right The result of "modulo". In one embodiment, as an example, let the optimal fault period be assumed. One sampling point, original vibration sample The length is 1024, for the sampling point number (Starting from 1 or 0, usually from 0), its corresponding phase group number ( (This is a modulo operation), then: Sampling point 0: Belongs to phase group 0; first sampling point: Phase group 1; ... 79th sampling point: Phase group 79; 80th sampling point: Then it returns to phase group 0; the 81st sampling point: phase group 1, and so on; thus, sampling points in the same relative position in all periods are assigned to the same phase group. For example, phase group 0 contains the 0th, 80th, 160th, ..., 960th sampling points (a total of 13 points), phase group 1 contains the 1st, 81st, 161st, ..., 961st sampling points, and so on.

[0041] (2) Calculate the phase energy for each phase group, and denote the phase energy as . , Represents phase group The degree of concentration of vibration and impact. In practical implementation, the phase group... The phase energy is obtained by averaging the absolute values ​​of all sampling points within the sample area. Optionally, to suppress the influence of a single outlier on the phase energy, a 3-point moving average smoothing can be performed on all phase energies in phase order. In one embodiment, following the previous example, phase group 0 contains multiple sampling points (e.g., 13), and the average of the absolute values ​​of these sampling points is calculated to obtain... Assuming the periodic impact in the original vibration sample is very strong, the average absolute value of phase group 0 (e.g., the phase in which the impact occurs) will be significantly higher than that of other phase groups; for example, , , , ..., .

[0042] To suppress individual outliers (such as a maximum value within a phase group that causes...) (Abnormally high), can be used to... arrive Performing a 3-point moving average, then: the new new ;in, This indicates the phase group number after the moving average processing. The smoothed phase energy, This indicates the phase group number after the moving average processing. The smoothed phase energy; the boundary can be mirrored or truncated, and the smoothed energy sequence can better reflect the overall trend between phase groups.

[0043] (3) Normalize all phase energies to the range of 0 to 1 to obtain the phase enhancement weights, which are denoted as . , Represents phase group The relative weights in localized impact enhancement. After normalization, phase groups with more concentrated periodic impacts correspond to larger phase enhancement weights, while background-dominated phase groups correspond to smaller phase enhancement weights. In one embodiment, as an example, all phase energies are... (or smoothed phase energy) Normalization to Range, assuming smoothing , , ..., the minimum value is 0.01, the maximum value is 0.25, then: ; ; That is, the phase group with the most concentrated periodic impacts gets the maximum weight of 1.0, while the weight of the background-dominated phase group is close to 0. In this way, during subsequent enhancement, the sampling points of phase group 0 will be significantly amplified, while phase groups 1, 2, etc. will hardly be enhanced.

[0044] (4) Determine the overall enhancement intensity based on the period matching score corresponding to the optimal candidate period, and denot the overall enhancement coefficient as follows: , This indicates that all phase enhancement weights are applied to the original vibration sample. The overall magnification is calculated as follows: ,in, This represents the period matching score corresponding to the optimal fault period.

[0045] Therefore, a higher cycle matching score indicates an optimal fault cycle. Higher reliability automatically increases the overall enhancement strength; conversely, lower cycle matching scores automatically decrease the overall enhancement strength. In practical implementation, the optimal failure cycle corresponds to the cycle matching score. It is calculated through step S201. Specifically, for each theoretical impact position, if an actual local maximum value is found within the allowable offset range, the amplitude is accumulated and finally divided by the number of theoretical positions. The larger the score, the more obvious and reliable the pattern of repeated impacts in the signal.

[0046] It should be noted that when When it is very large (e.g., much greater than 1), High overall strength; when When approaching 0, It hardly enhances; when hour, This allows the enhancement intensity to adapt to the periodic matching quality, avoiding blind amplification of random noise signals.

[0047] (5) For the original vibration sample By performing multiplicative enhancement point by point, impact-enhanced vibration samples are obtained. Impact-enhanced vibration sample This indicates the vibration sequence after periodic synchronous enhancement, with the size still being... In specific implementation, the first... The enhancement result of each sampling point is represented as: ,in, Represents the original vibration sample In the The amplitude at each sampling point Indicating impact-enhanced vibration samples In the The amplitude at each sampling point Indicates the first The phase group number to which each sampling point belongs.

[0048] It should be noted that, to avoid excessively large phase weights in some areas leading to abnormal amplification in certain regions, an upper limit is set for the enhancement factor, denoted as . , This represents the maximum amplification ratio that a single sampling point can achieve, for example, 1.8; if the calculated amplification factor of a sampling point exceeds... Then it is directly truncated to .

[0049] It should also be noted that the enhancement factor refers to the ratio of the enhanced amplitude to the original amplitude, i.e. This multiple may be greater than 1 or equal to 1 (when...). or To prevent individual sampling points from being over-amplified (e.g., magnified to more than 3 times the original value), an upper limit is set. That is, the maximum enhancement factor shall not exceed 1.8. If the calculated value of a certain sampling point is... If the value is 1.8, then the threshold is truncated. It should also be noted that although real fault impacts are weak, they usually have the characteristics of "recurring and relatively stable phases," while random noise spikes do not have this characteristic. This invention, through phase grouping and phase weighting, can preferentially enhance periodic impacts without introducing complex learning models. This not only improves the visibility of weak fault impacts but also avoids amplifying non-periodic noise as a whole, providing a more stable input for subsequent fault component separation.

[0050] In a specific implementation, step S3, based on the reconstruction of fault components and generation of statistical features according to singular spectral decomposition, is as follows: After S2 processing, the impact-enhanced vibration sample The periodic impacts in the sample are highlighted, but rotating fundamental frequencies, structural resonances, and broadband noise may still exist. If fixed-band filtering continues to be used, the effective components are easily weakened due to fault band shifts under different operating conditions. This invention uses singular spectrum decomposition to analyze the impact-enhanced vibration sample. By separating the dominant structure, reconstructed vibration samples that better characterize the fault impact patterns are output. And generate auxiliary statistical feature vectors. The specific steps are as follows: S3.1 Trajectory Matrix Construction and Principal Component Selection Impact-enhanced vibration sample The fault still contains rotating fundamental frequencies, structural resonances, and broadband noise. Fixed-band filtering is insufficient to adapt to fault band shifts, necessitating adaptive separation of repeating structures and background interference using a priori fault cycles. This invention utilizes the optimal fault cycle... Set embed length Constructing the trajectory matrix It then performs singular value decomposition and automatically determines the number of principal components to be retained based on the decay inflection point or cumulative energy threshold of the singular value sequence. The specific steps are as follows: (1) Based on the optimal failure cycle Set the embedding length, denoted as . , This represents the number of sampling points contained in each sliding window when constructing the trajectory matrix. In one implementation, let... This ensures that each sliding window covers at least approximately two fault impact cycles. Furthermore, to balance computational complexity and structural representation integrity, when... At that time, Revised to 64; when At that time, The cutoff value is 512.

[0051] (2) Vibration samples enhanced by impact As input, construct the trajectory matrix trajectory matrix Indicates that the vibration sample is enhanced by impact. According to length The two-dimensional matrix formed by the point-by-point expansion of the sliding window has a size of ,in, In practical implementation, from the impact-enhanced vibration sample The length of the cut-off point is 1. The continuous segments, as the trajectory matrix The first column; then move one sampling point forward and extract the next column of length. The continuous segments are used as the second column; this process is repeated until all end sampling points are covered. In one embodiment, as an example, suppose the impact-enhanced vibration sample... Length is Embedding length (Assuming) Then the trajectory matrix The size is ,in The construction process is as follows: Column 1: Take Sampling points 1-160 Column 2: Take the 2nd to 161st sampling points. Column 3: Take sampling points 3-162, ... Column 865: Take sampling points 865-1024. Each column is a sliding window segment of length 160, with 159 points overlapping between adjacent columns.

[0052] (3) For the trajectory matrix Performing singular value decomposition yields several singular values ​​and their corresponding singular components. The singular values ​​are denoted as... ,in, This represents the number of singular values, and the singular values ​​are sorted in descending order. The larger the singular value, the stronger the corresponding component's position in the trajectory matrix. The greater the contribution of the overall structure. Furthermore, the number of retained components is determined based on the magnitude of the singular value variation; the number of retained components is denoted as... , This indicates the number of dominant singular components used for fault component reconstruction. In the specific implementation, the difference between adjacent singular values ​​is first calculated, and then a significant decay inflection point is searched from front to back. When a certain difference is greater than both the previous and next differences, and this difference is greater than 1.5 times the average of all differences, the index corresponding to that position is taken as the number of retained components. If no obvious inflection point appears, the backup rule is adopted, that is, the smallest number of components whose cumulative singular value energy reaches 85% of the total energy is selected as the number of retained components. .

[0053] S3.2 Reconstruction of Vibration Samples and Generation of Statistical Features The dominant singular components are restored to a one-dimensional signal for use by the convolutional network, while retaining two complementary types of information: fault impact intensity and background noise stability. (Before stacking...) The reconstructed vibration sample is obtained by averaging the anti-angles after considering each singular component. , and by Obtain residual samples Then from and Eight statistical measures were extracted, including root mean square value, peak factor, impulse factor, and kurtosis, and concatenated into a statistical feature vector. The specific steps are as follows: (1) Before retention A dominant singular component, and will this The superposition of the dominant singular components yields an approximate trajectory matrix, which represents the impact-enhanced vibration sample. The matrix approximation results of the main repeating structures focus more on preserving the stable and repeating fault impact information. Furthermore, anti-angle averaging is performed on the approximate trajectory matrix to obtain the reconstructed vibration samples. Reconstructing vibration samples This represents a one-dimensional vibrational sequence recovered from the dominant singular components, with a size of... In practical implementation, the average of all elements in the approximate trajectory matrix located on the same anti-diagonal line is taken, and this average value is used as the reconstructed vibration sample. By sequentially iterating through all the anti-diagonals, the output value at the corresponding position in the middle can be used to reconstruct the vibration sample of the complete length. In one embodiment, as an example, assume an approximate trajectory matrix. Size is The anti-diagonal is defined as all lines that satisfy... The elements, among which, It is a row index (1-L). It is a column index (1-K), and the constant's value range is... arrive For example, constants Only elements The average value is the value of that element, corresponding to the reconstructed vibration sample. The first point; constant Occasionally there are elements and After averaging, we get The second point; constant From time to time After averaging, we get The third point; and so on, until the constant. Only elements ,get The 1 point (because the total number of anti-diagonal lines is ) (which is exactly equal to the original sample length); by traversing all anti-diagonal lines, a one-dimensional sequence can be recovered from a two-dimensional matrix. .

[0054] (2) Based on the impact-enhanced vibration sample With reconstructed vibration samples The difference is used to obtain the residual vibration sample. Residual vibration sample This represents the remaining portion retained by the non-dominant fault component, also with a size of [size missing]. , represented as Residual vibration sample It typically contains background noise, random perturbations, and unstable components that are not explained by the dominant component.

[0055] (3) From the reconstructed vibration samples respectively and residual vibration samples To extract time-domain statistics from the samples, and to balance fault impact intensity and background residual information, one implementation method involves reconstructing vibration samples... Calculate four statistics: root mean square value, peak factor, impulse factor, and kurtosis; for residual vibration samples... Also calculate the same four statistics; The root mean square (RMS) value is used to characterize the overall energy, the maxima factor is used to characterize the prominence of the peak value relative to the RMS value, the impulse factor is used to characterize the abrupt change of the peak value relative to the average absolute value, and the kurtosis is used to characterize the degree of peak sharpness. Furthermore, these eight statistics are concatenated in a fixed order to obtain a statistical feature vector. Statistical eigenvectors Indicates the reconstructed vibration samples and residual vibration samples The jointly formed auxiliary diagnostic features have a size of In a convenient implementation method, the arrangement order can be fixed as follows: reconstruct vibration samples The root mean square value, peak factor, impulse factor, kurtosis, and residual vibration samples. The root mean square value, peak factor, impulse factor, and kurtosis.

[0056] In one embodiment, as an example, if the vibration sample is reconstructed The root mean square value is 0.12, the peak factor is 3.8, the impulse factor is 4.6, and the kurtosis is 5.2; residual vibration samples If the root mean square value is 0.05, the peak factor is 2.1, the impulse factor is 2.4, and the kurtosis is 2.8, then the statistical eigenvector is... Can be written as This statistical feature vector It can simultaneously reflect the "strength of the dominant fault impact" and the "stability of the remaining background".

[0057] It should be noted that although the energy of early fault impacts is weak, they are usually somewhat repeatable, while background noise and random disturbances are more difficult to form a stable dominant structure in the trajectory matrix. This invention does not simply enhance the vibration samples with impact. Instead of overall noise reduction, it first utilizes the optimal fault cycle. After determining a more reasonable embedding length, singular spectral decomposition is used to separate repeating and non-repeating structures, resulting in output reconstructed vibration samples. It is more suitable for extracting stable fault modes in subsequent temporal convolutional branches, while statistical feature vectors It can provide supplementary information for judging the overall state.

[0058] In one embodiment, such as Figure 5 As shown, a comparison was made before and after periodic synchronous local impact enhancement to verify whether the multiplicative enhancement method proposed in this invention can effectively highlight periodic impacts. The experimental setup was as follows: based on the phase enhancement weight and the overall enhancement coefficient, the original vibration sample was multiplied point by point by the enhancement factor, with the upper limit of the enhancement factor set at 1.8 times. Figure 5The horizontal axis represents time (seconds), and the vertical axis represents amplitude (meters per square second). The blue curve represents the original fault signal, and the red curve represents the enhanced impact vibration sample. The comparison shows that at several time points where the impact occurred, the peak value of the red curve is significantly higher than that of the blue curve, indicating a moderate enhancement. In areas without impact, the two curves almost overlap, and background noise is not amplified. This indicates that the enhancement process preferentially amplifies periodically repeating impact structures, while having little impact on non-periodic random noise and fundamental frequency components.

[0059] In a specific implementation, step S4, the dual-input gating diagnostic network training and online early warning output, are as follows: Reconstructing vibration samples using only one-dimensional temporal convolution While it can extract local waveform patterns, its utilization of overall energy levels, residual fluctuation characteristics, and sample imbalance issues remains insufficient. This invention constructs a dual-input gated diagnostic network: [the network] reconstructs vibration samples... As a time series input, the statistical feature vector As auxiliary input, diagnostic features are obtained through gating compression after fusion, and then the prediction results of normal state and early fault state are output through class weighted training. The specific steps are as follows: S4.1 Forward Propagation of Timing Input Branch From reconstructed vibration samples Extracting local impact waveform patterns requires a lightweight structure to avoid overfitting and robustness to weak translation / scaling. This invention employs two one-dimensional convolutional layers: the first layer has 16 kernels (kernel length 32, stride 2), and the second layer has 32 kernels (kernel length 16, stride 1). Each layer is followed by an activation function and max pooling, and finally, global average pooling outputs a 32-dimensional temporal feature vector. The specific steps are as follows: (1) Reconstruct the vibration sample Input timing branch, reconstruct vibration sample It has a length of 1024 and contains the main fault impact structure that has been preserved after periodic enhancement and singular spectrum decomposition.

[0060] (2) Set up the first convolutional layer in the temporal input branch. The first convolutional layer uses 1D convolution operation, and the number of convolutional kernels is denoted as . , This represents the number of output channels of the first convolutional layer, for example, 16; the kernel length is denoted as... , This indicates the number of sampling points covered by each convolutional kernel, for example, 32; the convolution stride can be 2; the first convolutional layer is followed by a non-linear activation function and a max pooling layer to extract local impact patterns and reduce feature length.

[0061] (3) Set up a second convolutional layer in the temporal input branch, and denot the number of convolutional kernels in the second convolutional layer as follows: , This indicates the number of output channels of the second convolutional layer, for example, 32; the kernel length is denoted as... , The length of the kernel of the second convolutional layer can be 16, for example; the stride can be 1; the second convolutional layer is followed by a non-linear activation function and a max pooling layer to further extract stable local fault modes.

[0062] (4) Perform global average pooling on the multi-channel feature map output by the second convolutional layer to obtain the temporal feature vector. Temporal feature vector Indicates the reconstructed vibration sample The extracted temporal pattern features are denoted as follows: ,in, This represents the dimension of the time series feature, for example, it can be 32.

[0063] It should be noted that this invention employs a shallow 2-layer 1D convolutional structure in the temporal input branch, rather than a deep network or a complex recurrent structure, because of the single original vibration sample With fixed length and a typically limited number of early fault samples, shallow convolutions are more conducive to engineering implementation and training stability, while being sufficient to extract local patterns of the impact waveform.

[0064] S4.2, Statistical Input Branching, Feature Fusion and Gated Compression Directly concatenating heterogeneous features can introduce redundant or interfering dimensions, necessitating adaptive selection of important features. This invention uses statistical feature vectors. The fully connected mapping is used to obtain a 16-dimensional statistical feature vector. , and time series feature vector The concatenation yields a 48-dimensional fused feature vector. Then, after the first linear transformation, 64-dimensional intermediate features are obtained. Gated weights are generated in parallel and multiplied element by element. Finally, after the second linear transformation, a 32-dimensional diagnostic feature vector is obtained. The specific steps are as follows: (1) Statistical feature vector The statistical input branch is used to map the overall statistical information into a time-series feature vector. Auxiliary features of fusion.

[0065] (2) Set up a fully connected mapping layer in the statistical input branch to transfer the statistical feature vectors. Mapping from 8 dimensions to statistical feature vectors Statistical eigenvectors This represents the auxiliary diagnostic features extracted from the overall statistics, with dimensions denoted as . ,in, The dimension represents the statistical feature, for example, it can be 16; a normalization layer and a non-linear activation function can be connected after the fully connected mapping layer.

[0066] (3) Transform the time series feature vector With statistical eigenvectors Concatenate along dimensions to obtain the fused feature vector. fusion of feature vectors This indicates a combined feature that simultaneously includes local impact pattern information and overall statistical state information.

[0067] (4) For the fused feature vector Performing the first linear transformation yields an intermediate eigenvector, the dimension of which is denoted as . , This represents the intermediate representation dimension after the fused features are compressed, for example, 64; this linear transformation is followed by a nonlinear activation function.

[0068] (5) Generate gate weights based on the intermediate feature vectors, and weight the intermediate feature vectors dimension by dimension to obtain the gated feature vectors. Gated feature vectors This represents the fused features after adaptive filtering. In its implementation, a parallel fully connected layer outputs gate weights with the same dimension as the intermediate feature vector (this fully connected layer shares the fused feature vector with the linear transformation that generates the intermediate feature vector). As input, the feature vector is compressed to the range of 0 to 1 using the Sigmoid function, and then multiplied dimension by dimension with the intermediate feature vector to obtain the gated feature vector. Therefore, dimensions that contribute significantly will be retained, while dimensions with poor stability or low discriminative power will be weakened.

[0069] (6) Gated feature vectors Perform a second linear transformation to obtain the diagnostic feature vector. Diagnostic feature vector This represents the deep features ultimately used for classification, with dimensions denoted as . ,in, This indicates the dimension of the diagnostic features, for example, 32.

[0070] It should be noted that there are often only a few stable differences between weak early faults and normal states. Without gating screening, background fluctuations can easily interfere with the classification boundary. This invention does not reconstruct vibration samples. and statistical eigenvectors Instead of simply concatenating the data and sending it directly to the classification layer, we first extract the temporal pattern and statistical state separately, and then automatically determine "which feature dimensions should be retained and which feature dimensions should be suppressed" through gating weights, thereby reducing the impact of invalid features on the classification results and improving the discrimination stability of weak fault samples.

[0071] S4.3, Classification Output, Category-Weighted Training, and Online Early Warning Since there are far more normal samples than early fault samples, class imbalance needs to be mitigated. The probability output is then converted into an engineering early warning signal. In this invention, the output layer uses Softmax to obtain the probabilities of normal and early faults. During training, class weights are calculated based on the number of samples, and weighted cross-entropy loss is used for training constraints. The specific steps are as follows: (1) Diagnostic feature vector The input and output layers produce a 2D classification logic value vector, which corresponds to the unnormalized scores of the normal state and the early fault state, respectively. A Softmax transformation is then performed on these vectors to obtain the predicted probabilities of the normal state and the early fault state. The predicted probability of the early fault state is denoted as... The probability of predicting the normal state is denoted as .

[0072] (2) Calculate the class weights based on the number of normal samples and the number of early fault samples in the training set, where the number of normal samples is denoted as . The number of early failure samples is denoted as The normal category loss weight is denoted as The calculation method is expressed as The early failure category loss weight is denoted as The calculation method is expressed as Therefore, when the number of early faulty samples is less than the number of normal samples, It will automatically be greater than This is to increase the model's attention to early failure samples.

[0073] (3) Calculate the weighted cross-entropy loss based on the class weights, and use this loss to train the dual-input gated diagnostic network via backpropagation. The Adam optimizer can be used for training, with an initial learning rate of 0.001, a batch size of 64, and a weight decay coefficient of [missing value]. During training, the stopping point can be determined by the validation set loss or the validation set F1 score. It should be noted that the method of calculating the weighted cross-entropy loss by class weights assigns different weights to different classes when calculating the cross-entropy loss, in order to mitigate the bias effect caused by the imbalanced class distribution in the training samples. This method is commonly used in machine learning to handle class imbalance problems.

[0074] (4) After completing the training, new original vibration samples are used. Execute S2 and S3 sequentially to obtain reconstructed vibration samples. and statistical eigenvectors Then reconstruct the vibration samples The input time series input branch will generate statistical feature vectors. Input statistics branch, output early fault state prediction probability .

[0075] (5) Predict the probability of early failure states The warning result is output after comparing it with the preset warning threshold. The warning threshold is denoted as [missing information]. , This represents the minimum probability threshold for triggering an early fault warning, for example, 0.5; in scenarios that place greater emphasis on early recall, it can be... Lowered to between 0.4 and 0.5, that is: when When, output early fault warning results; when When the time is right, output the normal state result.

[0076] It should be noted that the dual-input gated diagnostic network consists of two parallel branches and a fusion gate module. Its core feature is to use a gating mechanism to adaptively select important dimensions in the fused features and suppress irrelevant or interfering features. The structure is as follows: a) Timing input branch: Receives reconstructed vibration samples (Length 1024), contains two one-dimensional convolutional layers, each followed by a non-linear activation function and a max-pooling layer. Specifically, First convolutional layer: 16 convolutional kernels, kernel length 32, stride 2; Second convolutional layer: 32 convolutional kernels, kernel length 16, stride 1; Finally, the temporal feature vector is output through global average pooling. (Dimension 32).

[0077] b) Statistical input branch: receives statistical feature vectors (8-dimensional), containing a fully connected layer that maps 8 dimensions to 16 dimensions, followed by a normalization layer and a ReLU activation function, outputting a statistical feature vector. (Dimension 16).

[0078] c) Feature fusion and gated compression: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] and By splicing along the dimensions, the fused features are obtained. (dimension) First linear transformation: The fully connected layer will Mapping to intermediate feature vectors (Dimension 64), followed by a ReLU activation function; parallel gating generation: the output of another fully connected layer and Gated weight vectors of the same dimension are compressed to the (0,1) range by the Sigmoid activation function, and then... Element-wise multiplication yields the gating feature. Second linear transformation: The fully connected layer will Mapping to diagnostic feature vectors (Dimension 32).

[0079] d) Output layer: The fully connected layer will Mapped to 2D, and then processed by Softmax to output the probabilities of normal operation and early failure.

[0080] In a specific implementation, step S5 is as follows: After model training is complete, this invention deploys the trained dual-input gated diagnostic network in the online monitoring system for fan coil units. In practical applications, the sampling frequency is consistent with that used during the training phase. Vibration signals are acquired in real time and divided into original vibration samples of a fixed length of 1024 sampling points. Then, for each original vibration sample to be tested... The optimal fault period determination in S2.1 and the periodic synchronous local impact enhancement in S2.2 are executed sequentially to obtain impact-enhanced vibration samples. Then, execute steps S3.1 (trajectory matrix construction and principal component selection) and S3.2 (reconstruction of vibration samples and generation of statistical features) to obtain the reconstructed vibration samples. and statistical eigenvectors Finally, the vibration samples will be reconstructed. and statistical eigenvectors The dual-input gating diagnostic network, trained with common inputs, outputs the predicted probability of normal state and the predicted probability of early fault state, and gives the final warning result based on the preset warning threshold.

[0081] It should be noted that this invention forms a complete closed-loop processing flow of "periodic prior enhancement - fault component reconstruction - dual-input gating discrimination". Periodic prior enhancement is used to improve the visibility of weak, repetitive impacts; fault component reconstruction is used to separate stable fault structures from background disturbances; and dual-input gating discrimination is used to jointly utilize local waveform patterns and overall statistical states. Compared with schemes that only use fixed filtering or ordinary classification networks, this method is more suitable for early-stage fault scenarios of low-speed, weak-impact, and strong-interference fan coil motors. It can improve the sensitivity of weak fault identification and classification stability while maintaining a clear implementation path and engineering feasibility.

[0082] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for early warning and diagnosis of latent faults in fan coil unit motors, characterized in that, during operation... as follows: Vibration signals from fan coil unit motors are collected and organized into training samples. The operating status of each sample is labeled to form a set of labeled training samples. The optimal fault period is determined by utilizing the repetitive impact pattern of training samples in the time domain. Synchronous reinforcement is then performed on the same phase position according to the optimal fault period to obtain impact-enhanced vibration samples. Singular spectral decomposition is used to separate the dominant structure of the impact-enhanced vibration sample, and the reconstructed vibration sample and statistical feature vector are output. The reconstructed vibration samples are used as the main input and the statistical feature vectors are used as the auxiliary input to the constructed dual-input gated diagnostic network. After fusion, the diagnostic feature vectors are obtained through gating compression. Then, the prediction results of normal state and early fault state are output through class weighted training. The trained dual-input gated diagnostic network is deployed in the online monitoring system of fan coil motor to monitor the status of fan coil motor in real time.

2. The method for early warning and diagnosis of latent faults in fan coil unit motors according to claim 1, characterized in that, The process of determining the optimal failure period is as follows: The absolute values ​​of the training samples are taken point by point to obtain an absolute value vibration sequence. The location of local maxima is detected in the absolute value vibration sequence, based on the equipment rotation speed range and sampling frequency. Set a candidate period range, select a reference impact position from the local maximum positions, and for each theoretical impact position, if an actual local maximum position is found within the allowable offset range, record the absolute value amplitude corresponding to the actual local maximum position. After traversing all theoretical impact positions, divide the cumulative amplitude by the number of theoretical impact positions to obtain the period matching score corresponding to the current candidate period. Calculate the period matching score for each candidate period within the candidate period range, and select the candidate period with the largest period matching score from all candidate periods as the optimal fault period.

3. The method for early warning and diagnosis of latent faults in fan coil unit motors according to claim 1, characterized in that, The calculation process for the impact-enhanced vibration sample is as follows: The optimal fault cycle is moduloed according to the sampling point sequence number, and the training samples are divided into phase groups. Each phase group represents the relative position of the sampling point in one cycle, and the phase group number to which each sampling point belongs is its modulo of the optimal fault cycle. Then, the phase energy of each phase group is statistically analyzed, and the absolute values ​​of all sampling points within the phase group are averaged to obtain the phase energy. Then, all phase energies are normalized to the range of 0 to 1 to obtain the phase enhancement weight. Then, the overall enhancement intensity is determined according to the period matching score corresponding to the optimal candidate period. The overall enhancement intensity represents the overall amplification degree of all phase enhancement weights applied to the training samples. Finally, multiplicative enhancement is performed on the training samples point by point to obtain the impact-enhanced vibration samples.

4. The method for early warning and diagnosis of latent faults in fan coil unit motors according to claim 1, characterized in that, The embedding length is set according to the optimal fault cycle, the trajectory matrix is ​​constructed and singular value decomposition is performed, and then the number of principal components to be retained is automatically determined by the decay inflection point or cumulative energy threshold of the singular value sequence. Using impact-enhanced vibration samples as input, a trajectory matrix is ​​constructed. The trajectory matrix represents the trajectory of the impact-enhanced vibration samples along a length of... The two-dimensional matrix is ​​formed by expanding the sliding window point by point. Indicates the embedding length; Singular value decomposition is performed on the trajectory matrix to obtain several singular values ​​and corresponding singular components. The singular values ​​are sorted from largest to smallest. The larger the singular value, the greater the contribution of the corresponding component to the overall structure of the trajectory matrix. Then, the number of components to be retained is determined based on the magnitude of the change in singular values, and then the number of dominant singular components to be retained is determined.

5. The method for early warning and diagnosis of latent faults in fan coil unit motors according to claim 4, characterized in that, The retained dominant singular components are superimposed on the impact-enhanced vibration samples to obtain reconstructed vibration samples. Then, the statistical feature vector is determined based on the residual between the enhanced vibration samples and the reconstructed vibration samples. The retained dominant singular components are superimposed on the impact-enhanced vibration samples to obtain an approximate trajectory matrix. Then, the approximate trajectory matrix is ​​averaged in the opposite direction to obtain a reconstructed vibration sample. Finally, the residual vibration sample is obtained based on the difference between the impact-enhanced vibration sample and the reconstructed vibration sample. The residual vibration sample contains background noise, random disturbances, and unstable components not explained by the dominant components. Time-domain statistics are extracted from the reconstructed vibration samples and the residual vibration samples respectively. The root mean square value, peak factor, impulse factor and kurtosis are calculated for the reconstructed vibration samples and the residual vibration samples respectively. The eight statistics are concatenated in a fixed order to obtain the statistical feature vector.

6. The method for early warning and diagnosis of latent faults in fan coil unit motors according to claim 1, characterized in that, The dual-input gating diagnostic network includes a timing input branch and a statistical input branch; The structure of the timing input branch and the forward propagation process are as follows: The temporal input branch has two layers of one-dimensional convolutions: the first layer has 16 kernels, and the second layer has 32 kernels. Each layer is followed by an activation function and max pooling. Finally, global average pooling outputs a 32-dimensional temporal feature vector. The reconstructed vibration sample is input into the temporal input branch. The reconstructed vibration sample contains the main fault impact structure retained after periodic enhancement and singular spectrum decomposition. It is then passed through the first convolutional layer and the second convolutional layer in sequence. Global average pooling is performed on the multi-channel feature map output by the second convolutional layer to obtain the temporal feature vector.

7. The method for early warning and diagnosis of latent faults in fan coil unit motors according to claim 6, characterized in that, The statistical input branch contains a fully connected mapping layer, which can be followed by a normalization layer and a non-linear activation function. The statistical feature vector is input into the statistical input branch and mapped to a statistical feature vector.

8. The method for early warning and diagnosis of latent faults in fan coil unit motors according to claim 7, characterized in that, The temporal feature vector and the statistical feature vector are concatenated by dimension to obtain the fused feature vector. The first linear transformation is performed on the fused feature vector to obtain the intermediate feature vector. After the linear transformation, a nonlinear activation function is applied to generate gating weights based on the intermediate feature vector. The intermediate feature vector is then weighted dimension by dimension to obtain the gated feature vector. The second linear transformation is performed on the gated feature vector to obtain the diagnostic feature vector.

9. The method for early warning and diagnosis of latent faults in fan coil unit motors according to claim 1, characterized in that, The diagnostic feature vector is processed using Softmax to obtain the probabilities of normal and early failure. During training, class weights are calculated based on the number of samples, and weighted cross-entropy loss is used for training constraints. The specific operation is as follows: The diagnostic feature vector is input into the output layer to obtain a 2D classification logic value vector, which corresponds to the unnormalized scores of the normal state and the early fault state, respectively. Then, a Softmax transformation is performed on it to obtain the predicted probabilities of the normal state and the early fault state. The class weights are calculated based on the number of normal samples and the number of early fault samples in the training sample set. The weighted cross-entropy loss is calculated based on the class weights, and the dual-input gating diagnostic network is trained by backpropagation using this loss. After training, new sample data is input into the trained network, and the early fault state prediction probability is output. The early fault state prediction probability is compared with the preset warning threshold, and the warning result is output.

10. The method for early warning and diagnosis of latent faults in fan coil unit motors according to claim 1, characterized in that, The training sample set contains all training samples and their corresponding labels, as follows: Using vibration sensors Vibration signals from the fan coil motor housing or bearing housing are collected at a sampling frequency. The continuously collected vibration signals are divided into multiple training samples by a fixed length. Each training sample represents a continuous vibration amplitude sequence obtained from one segmentation. Each training sample is labeled, and the label is denoted as . , This indicates the running status of the training samples. Indicates a normal state. This indicates an early stage of a fault.

Images