CN114838496B - Air conditioner muffler performance detection method based on artificial intelligence - Google Patents
Air conditioner muffler performance detection method based on artificial intelligence Download PDFInfo
- Publication number
- CN114838496B CN114838496B CN202210432016.4A CN202210432016A CN114838496B CN 114838496 B CN114838496 B CN 114838496B CN 202210432016 A CN202210432016 A CN 202210432016A CN 114838496 B CN114838496 B CN 114838496B
- Authority
- CN
- China
- Prior art keywords
- muffler
- sample
- training set
- network
- target training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000013473 artificial intelligence Methods 0.000 title claims abstract description 14
- 238000001514 detection method Methods 0.000 title claims abstract description 12
- 238000012549 training Methods 0.000 claims abstract description 199
- 230000006870 function Effects 0.000 claims abstract description 78
- 230000003584 silencer Effects 0.000 claims abstract description 60
- 238000000034 method Methods 0.000 claims abstract description 59
- 230000008569 process Effects 0.000 claims abstract description 36
- 239000013598 vector Substances 0.000 claims description 150
- 210000002569 neuron Anatomy 0.000 claims description 45
- 238000002372 labelling Methods 0.000 claims description 32
- 238000004378 air conditioning Methods 0.000 claims description 5
- 238000012360 testing method Methods 0.000 claims 1
- 230000005236 sound signal Effects 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 4
- 238000009434 installation Methods 0.000 description 4
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000030279 gene silencing Effects 0.000 description 2
- 238000011478 gradient descent method Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
- F24—HEATING; RANGES; VENTILATING
- F24F—AIR-CONDITIONING; AIR-HUMIDIFICATION; VENTILATION; USE OF AIR CURRENTS FOR SCREENING
- F24F13/00—Details common to, or for air-conditioning, air-humidification, ventilation or use of air currents for screening
- F24F13/24—Means for preventing or suppressing noise
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Abstract
The invention relates to an air conditioner muffler performance detection method based on artificial intelligence, and belongs to the technical field of muffler performance detection. The method comprises the following steps: acquiring a target training set; performing network training based on the target training set to obtain a target network; acquiring data to be detected; and inputting the data to be detected into a target network to obtain the performance grade of the automobile air conditioner silencer corresponding to the data to be detected. The invention adjusts the network learning rate based on the loss function value of the network in the multi-round training process and the network parameters during training, can train the network efficiently and precisely, and can detect the performance of the automobile air conditioner silencer relatively reliably based on the target network.
Description
Technical Field
The invention relates to the technical field of muffler performance detection, in particular to an air conditioner muffler performance detection method based on artificial intelligence.
Background
Along with the development of social economy and the improvement of life of people, the comfort of people on the use of automobiles is also improved, namely, the use performance requirement of the automobile air-conditioning silencer is also improved, so that the performance detection of the automobile air-conditioning silencer in the use process is very important, wherein the automobile air-conditioning silencer is used for reducing the machine noise of an air conditioner in the use process.
The existing method for detecting the performance of the automobile air conditioner silencer in the using process based on manpower is strong in subjectivity and low in detection efficiency, and can be found only when a large problem occurs in the using process of the automobile air conditioner silencer, so that the method for detecting the performance of the automobile air conditioner silencer based on manpower is low in reliability.
Disclosure of Invention
The invention provides an artificial intelligence-based air conditioner muffler performance detection method, which is used for solving the problem that the performance of an automobile air conditioner muffler cannot be reliably detected in the existing method, and adopts the following technical scheme:
in a first aspect, an embodiment of the present invention provides an artificial intelligence based method for detecting performance of an air conditioner muffler, including the steps of:
acquiring a target training set;
performing network training based on the fixed network learning rate and the target training set to obtain a first loss function value sequence corresponding to the first network and the target training set; judging whether the stability degree of the first loss function value sequence is smaller than a preset stability degree threshold value, if so, obtaining a first target learning rate of each data in a target training set according to the first loss function value sequence; training the first network based on the first target learning rate and the target training set to obtain a second network and a second loss function value sequence corresponding to the target training set; judging whether the stability degree of the second loss function value sequence is smaller than a preset stability degree threshold value or not, if so, obtaining a second target learning rate of each data in a target training set according to the second loss function value sequence; training the second network based on the second target learning rate and the target training set to obtain a third network and a third loss function value sequence corresponding to the target training set; and so on; stopping network training until the stability degree of the loss function value sequence is greater than or equal to a preset stability degree threshold value, and obtaining a target network;
acquiring data to be detected; and inputting the data to be detected into a target network to obtain the performance grade of the automobile air conditioner silencer corresponding to the data to be detected.
The beneficial effects are that: the invention takes the target training set as the basis for obtaining the target network; taking the target network as a basis for obtaining the performance grade of the automobile air conditioner silencer; the invention adjusts the network learning rate based on the loss function value of the network in the multi-round training process and the network parameters during training, can train the network efficiently and precisely, and can detect the performance of the automobile air conditioner silencer relatively reliably based on the target network.
Preferably, the target training set includes a plurality of sample muffler feature vectors.
Preferably, network training is performed based on a fixed network learning rate and a target training set to obtain a first loss function value sequence corresponding to the first network and the target training set; judging whether the stability degree of the first loss function value sequence is smaller than a preset stability degree threshold value, if so, obtaining a first target learning rate of each data in a target training set according to the first loss function value sequence, wherein the method comprises the following steps:
performing performance grade labeling on each sample muffler feature vector in the target training set;
training based on the characteristic vectors of the various sample silencers and the labeling performance grades of the characteristic vectors of the various sample silencers to obtain a first network;
calculating to obtain squares of differences between marking performance grades corresponding to the characteristic vectors of the sample mufflers and network prediction performance grades in the first network process, and obtaining first loss function values of the characteristic vectors of the sample mufflers;
constructing and obtaining a first loss function value sequence corresponding to the target training set according to the first loss function value;
calculating the standard deviation of the first loss function value sequence, and recording the inverse of the standard deviation of the first loss function value sequence as the stability of the first loss function value sequence;
and judging whether the stability degree of the first loss function value sequence is smaller than a preset stability degree threshold value, and if so, obtaining a first target learning rate of each sample muffler characteristic vector in a target training set according to the first loss function value sequence.
Preferably, the method for obtaining the first target learning rate of the characteristic vector of each sample muffler in the target training set according to the first loss function value sequence includes:
obtaining feature vectors of each sample muffler corresponding to each labeling performance grade in the target training set;
obtaining a first difference value of each sample muffler feature vector in the target training set according to the Euclidean distance between any two sample muffler feature vectors corresponding to each labeling performance grade;
obtaining a second difference value of each sample muffler feature vector in the target training set according to the average value of the sum of Euclidean distances between each sample muffler feature vector corresponding to each labeling performance level and each sample muffler feature vector corresponding to the rest labeling performance levels;
obtaining the performance grade distinguishing difficulty corresponding to each sample muffler feature vector according to the ratio of the first difference value of each sample muffler feature vector to the second difference value of the corresponding sample muffler feature vector in the target training set;
obtaining a first reference learning rate corresponding to a target training set according to the first loss function value of each sample muffler characteristic vector and the performance grade distinguishing difficulty of each sample muffler characteristic vector; obtaining a first reference learning rate of each sample muffler feature vector according to the first reference learning rate corresponding to the target training set;
obtaining a first output parameter value of each neuron corresponding to each sample silencer feature vector in the target training set and a first output parameter value sequence of each neuron corresponding to the target training set according to the process of obtaining the first network by the target training set;
obtaining a process of a first network and an inverse feedback network according to the target training set, and obtaining first parameter adjustment gradients of neurons corresponding to the characteristic vectors of the silencer of each sample in the target training set;
obtaining the importance degree of each neuron in a first network according to the absolute value of the partial correlation coefficient between the first output parameter value sequence and the corresponding first loss function value sequence;
according to the importance degree of each neuron and the first parameter of each neuron corresponding to each sample silencer feature vector, adjusting the gradient to obtain an adjustment coefficient of a first reference learning rate corresponding to each sample silencer feature vector;
and obtaining a first target learning rate of each sample muffler feature vector in the target training set according to the adjustment coefficient of the first reference learning rate corresponding to each sample muffler feature vector and the first reference learning rate corresponding to each sample muffler feature vector.
Preferably, the first variance value of each sample muffler feature vector in the target training set is calculated according to the following formula:
wherein B is i For the first difference value of the ith sample muffler characteristic vector corresponding to a certain labeling performance level in the target training set, B1 i,j And I is the number of the sample muffler feature vectors in the labeling performance grade in the target training set, wherein the Euclidean distance between the ith sample muffler feature vector corresponding to the labeling performance grade in the target training set and the jth sample muffler feature vector corresponding to the labeling performance grade is obtained.
Preferably, the first reference learning rate corresponding to the target training set is calculated according to the following formula:
wherein C is a first reference learning rate corresponding to the target training set, delta is a super parameter, b1 is the number of sample muffler feature vectors in the target training set, and C1 b Difficulty, C2 for distinguishing performance grade of b sample muffler characteristic vector in target training set b A first loss function value for the b-th sample muffler feature vector in the target training set.
Preferably, the method for calculating the adjustment coefficient of the first reference learning rate corresponding to each sample muffler feature vector according to the following formula includes:
wherein F is an adjustment coefficient of a first reference learning rate corresponding to a certain sample muffler feature vector in the target training set, M is the number of neurons in a first network, Y m G is the importance of the mth neuron in the first network m And (3) adjusting the gradient for the first parameter of the m-th neuron corresponding to the silencer feature vector in the target training set, wherein τ is a super parameter.
Drawings
FIG. 1 is a flow chart of an artificial intelligence based method for detecting the performance of an air conditioner muffler.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is apparent that the described embodiments are only some embodiments of the present invention, but not all embodiments, and all other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the embodiments of the present invention.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
The embodiment provides an air conditioner muffler performance detection method based on artificial intelligence, which is described in detail as follows:
as shown in fig. 1, the method for detecting the performance of the muffler of the air conditioner based on the artificial intelligence comprises the following steps:
step S001, obtaining a target training set.
The performance of the automobile air conditioner silencer is detected based on the network, so that the network is trained, and a reasonable network learning rate is required to be set in order to ensure high-efficiency and high-precision network training; when the learning rate setting is too small, the convergence process becomes very slow, and when the learning rate setting is too large, learning is accelerated in the early stage of network training, so that the model is easier to approach to a local or global optimal solution, but in the later stage, larger fluctuation exists, even the value of the loss function loiters around the minimum value, and the convergence cannot be realized. Therefore, the network learning rate is adjusted according to the error condition of the network in the training process and the network parameters during training, so that the network can be trained efficiently and precisely, and the performance of the automobile air conditioner silencer can be detected relatively reliably based on the trained network.
The performance of the silencer of the automobile air conditioner is mainly evaluated by the acoustic performance, the flow field performance and the mechanical performance of the silencer; the sound performance is evaluated by the sound-deadening effect of the muffler, so the present embodiment collects sound signals at the near-source end, which is the position where the muffler has not deadened, and the far-source end, which is the position where the muffler has deadened, respectively, and the device for collecting sound signals may be a sensor or a recording device; the advantages and disadvantages of the flow field performance are evaluated by the influence of the silencer on air resistance, the air conditioner needs smoother air circulation, the air conditioner silencer is used as a part of the air conditioner, if the flow field performance is poor, the air conditioner system cannot normally operate or the working efficiency of the air conditioner can be influenced, the flow field performance is generally reflected by resistance loss or resistance coefficient, the resistance loss can be reflected by air flow, therefore, the air flow rate is collected at the air inflow position of the silencer, the air flow rate is collected at the air outflow position of the silencer, and the air flow rate can be collected by a sensor; the mechanical performance is evaluated mainly by indexes such as the size, the service life and the price of the silencer, the performance is better when the size is smaller, the service life is longer and the price is lower, the size mainly influences the installation condition of the silencer, namely the installation efficiency is influenced when the size of the silencer is too large, the installation of other parts is interfered, the service life of the silencer is the average service life of various silencers, and the price of the silencer is the market price of the silencer.
In this embodiment, the performances of the three aspects are both related and restricted, and the larger the silencing amount is, the better the silencing performance of the silencer is, but the requirements of the flow field performance must be considered, the acoustic performance and the flow field performance must be considered, and the requirements of the mechanical performance must be considered, so that the silencer is economical and durable, and the problems of overlarge volume, difficult installation and the like are avoided.
Therefore, the embodiment can acquire the near-source end sound signal data, the far-source end sound signal data, the muffler air inflow position air flow rate, the muffler air outflow position air flow rate, the muffler volume, the muffler life and the muffler market price corresponding to each running time of the automobile air conditioner muffler through the data acquisition method, and construct and acquire each sample muffler feature vector a= { A1, A2, A3, A4, A5, A6 and A7}, wherein A1 is the near-source end sound signal data corresponding to any running time of the automobile air conditioner muffler, A2 is the far-source end sound signal data corresponding to any running time of the automobile air conditioner muffler, A3 is the muffler air inflow position air flow rate corresponding to any running time of the automobile air conditioner muffler, A4 is the muffler air outflow position air flow rate corresponding to any running time of the automobile air conditioner muffler, A5 is the muffler volume, A6 is the muffler life and A7 is the muffler market price; all information contained in the one sample muffler feature vector is obtained based on the same muffler run time.
Because the performance of the silencer is detected through the network, a plurality of sample silencer feature vectors of the air-conditioner silencers in the running process of the plurality of automobiles can be obtained through the data acquisition process, and the plurality of sample silencer feature vectors are recorded as a target training set; the number of the characteristic vectors of the sample muffler in the target training set is set to 3000; and the muffler type to which the sample muffler feature vector corresponds is intended to include all muffler types that occur on the market. As other embodiments, the number of sample muffler feature vectors acquired by other numbers, such as 6000, etc., may also be set.
And step S002, performing network training based on the target training set to obtain a target network.
In the embodiment, the performance grade of the automobile air conditioner silencer is detected through the neural network, and in order to reduce the workload and enhance the network accuracy, the target training set is required to be reused for network training, so that a target network is obtained, the target network is the basis for subsequently detecting the performance grade of the automobile air conditioner silencer, and the learning degree of the target network is deeper; the specific process for obtaining the target network comprises the following steps:
because the target training set can completely learn the information of the target training set without training once in the network training process, the target training set is required to be trained repeatedly for a plurality of times under the general condition, namely, the target training set is trained for a plurality of times, and different learning rates are required to be set when the target training set is trained for a plurality of times in order to ensure the network convergence and the learning degree of the network; therefore, the embodiment sets the learning rate of the next round of network based on the learning degree of the previous round of network, and the specific process is as follows:
in the embodiment, an expert in the aspect of setting silencer performance evaluation marks the performance level of each sample silencer characteristic vector in a target training set, and then performs a first round of network training based on the target training set and the marked performance level of each sample silencer characteristic vector in the target training set to obtain a first network; the network learning rate is fixed and higher in the process of training to obtain the first network; the loss function is a mean square error loss, the structure of the network is an Encoder-FC structure, and the specific network structure and training process of the network are the prior art, so this embodiment will not be described in detail.
Because the network can learn fully through multiple rounds generally, but in order to train the target training set blindly, the embodiment sets the cut-off condition of the network training again on the basis of determining the learning rate of the next round of network based on the learning degree of the previous round of network; calculating and training to obtain the square of the difference between the labeling performance grade corresponding to each sample muffler characteristic vector in the first network process and the network prediction performance grade in the training process, and obtaining a first loss function value of each sample muffler characteristic vector; constructing and obtaining a first loss function value sequence corresponding to the target training set according to the first loss function values of the characteristic vectors of the silencer of each sample; calculating the standard deviation of the first loss function value sequence, and recording the inverse number of the standard deviation corresponding to the first loss function value sequence as the stability degree of the first loss function value sequence, wherein when the stability degree is larger, the first network learning is more sufficient, and otherwise, the first network learning is insufficient; judging whether the stability degree of the first loss function value sequence is smaller than a preset stability degree threshold value or not, if so, obtaining a first target learning rate of each silencer feature vector in a target training set according to the learning degree of a first network; the better the learning degree of the first network is, the more sufficient the first network is learned, and the higher the network precision is; the stability of the loss function value sequence is larger than a preset stability threshold value after a plurality of times of training due to the characteristics of the neural network training process; the preset stability threshold value needs to be set according to actual conditions.
As another embodiment, the first network may be a network after training for five or three rounds based on the target training set and the labeling performance level of each sample muffler feature vector in the target training set.
According to the learning degree of the first network, the specific process of obtaining the first target learning rate of each muffler feature vector in the target training set is as follows:
because when the performance grade distinguishing difficulty of the characteristic vectors of the sample silencers is small, relatively accurate grade distinguishing can be realized only by a network with shallow learning degree, and when the performance grade distinguishing difficulty of the characteristic vectors of the silencers is large, relatively accurate grade distinguishing can be realized only by a network with good learning degree, therefore, if the reliability of the first target learning rate of the characteristic vectors of the silencer of each sample in the target training set is obtained only based on the first loss function values corresponding to the characteristic vectors of the silencer of each sample in the target training set is low.
Therefore, in this embodiment, the difficulty of distinguishing the performance levels corresponding to the feature vectors of the muffler of each sample is calculated according to the distance between the feature vectors of the muffler of each sample; the specific process for obtaining the performance grade distinguishing difficulty corresponding to the characteristic vector of each sample muffler comprises the following steps:
obtaining feature vectors of each sample muffler corresponding to each labeling performance grade in the target training set; obtaining a first difference value of each sample muffler feature vector corresponding to each labeling performance grade in the target training set according to the Euclidean distance between any two sample muffler feature vectors corresponding to each labeling performance grade; calculating a first difference value of each sample muffler characteristic vector corresponding to each labeling performance grade in the target training set according to the following formula:
wherein B is i For the first difference value of the ith sample muffler characteristic vector corresponding to a certain labeling performance level in the target training set, B1 i,j And I is the number of the sample muffler feature vectors in the labeling performance grade in the target training set, wherein the Euclidean distance between the ith sample muffler feature vector corresponding to the labeling performance grade in the target training set and the jth sample muffler feature vector corresponding to the labeling performance grade is obtained.
And calculating the sum of Euclidean distances between the ith sample muffler feature vector corresponding to a certain labeling performance level in the target training set and each sample muffler feature vector corresponding to each labeling performance level except the labeling performance level in the corresponding target training set, and averaging the sum of Euclidean distances to obtain a second difference value of the ith sample muffler feature vector corresponding to the labeling performance level in the target training set, namely obtaining the second difference value of each sample muffler feature vector corresponding to each labeling performance level in the target training set.
Therefore, the first difference value and the second difference value corresponding to the characteristic vector of each sample muffler in the target training set can be obtained through the process; calculating the ratio of a first difference value corresponding to each sample muffler feature vector in the target training set to a corresponding second difference value to obtain the performance grade distinguishing difficulty corresponding to each sample muffler feature vector, and when the first difference value corresponding to a certain sample muffler feature vector is smaller and the second difference value is larger, indicating that the performance grade distinguishing difficulty of the sample muffler feature vector is smaller, otherwise, the performance grade distinguishing difficulty is larger; and normalizing the performance grade distinguishing difficulty.
Obtaining a first reference learning rate corresponding to each sample muffler feature vector according to the first loss function value of each sample muffler feature vector and the difficulty in distinguishing the performance grade of each normalized sample muffler feature vector; calculating a first reference learning rate corresponding to the target training set according to the following formula:
wherein C is a first reference learning rate corresponding to the target training set, delta is a super parameter, b1 is the number of sample muffler feature vectors in the target training set, and C1 b Difficulty, C2 for distinguishing performance grade of b sample muffler characteristic vector in target training set b A first loss function value for a b-th sample muffler feature vector in the target training set; the larger C indicates that the learning rate is higher when the first network is trained based on the target training set, and the larger C indicates that the learning degree of the first network is shallower, and the higher network learning rate is required to be set when the first network is optimized based on the target training set later; and the first reference learning rate corresponding to the target training set is also the first reference learning rate of each sample muffler feature vector when the first network is optimized based on the target training set.
Because the learning requirement of the network on each data is different, the first reference learning rate of each obtained sample muffler feature vector is mainly based on the first loss function value without referring to the influence of the target training set on the network parameters in the process of training to obtain the first network, the first reference learning rate corresponding to the target training set needs to be adjusted according to the output parameter value of each neuron in the network in the process of obtaining the first network by combining the target training set, and the important condition of each sample muffler feature vector in the target training set is considered when the adjustment is performed, when the change of the important parameter in the network is larger due to each sample muffler feature vector in the target training set, the first reference learning rate adjustment coefficient of each sample muffler feature vector can be determined based on the output parameter value of each neuron in the process of obtaining the first network and the important condition of each sample muffler feature vector in the target training set when the target training set optimizes the first network; the specific process is as follows:
in this embodiment, in the process of obtaining the first network based on the target training set, after obtaining the first loss function value of each sample muffler feature vector, the network will update the parameter sizes of all neurons in the neural network by using a random gradient descent method, so that the output parameter values of each neuron corresponding to each sample muffler feature vector in the process of obtaining the target training set to obtain the first network are required; the first parameter adjustment gradient of each neuron corresponding to each sample muffler feature vector in the process of obtaining the target training set through the inverse feedback network is obtained, and the larger the parameter adjustment gradient is, the poorer the description precision of the parameter to the sample muffler feature vector is explained, so that the requirement of the neuron on the learning of the data is higher; for example, after obtaining the first loss function value of a certain sample muffler feature vector, the neural network updates the parameter sizes of all neurons in the network by using a random gradient descent method, that is, the parameters of each neuron in the network are correspondingly adjusted, so as to obtain output parameter values corresponding to each neuron, and the output parameter values corresponding to each neuron are recorded as the first output parameter values of each neuron corresponding to the sample muffler feature vector; therefore, in the process of obtaining the first network through the target training set, the first output parameter value of each neuron corresponding to each sample silencer feature vector can be obtained, and further, the first output parameter value sequence of each neuron corresponding to the target training set is obtained.
Obtaining absolute values of partial correlation coefficients between the first output parameter value sequences of the neurons corresponding to the target training set and the corresponding first loss function value sequences, wherein the partial correlation coefficients can reflect correlation between the first output parameter value sequences of the neurons and the first loss function value sequences, and the larger the partial correlation coefficients are, the larger the correlation degree between the corresponding neurons and the first loss function values is, namely, the larger the neuron parameter value changes influence the classification precision of a network, so that when the partial correlation coefficients between the first output parameter value sequences of the neurons and the first loss function value sequences are larger, the more important the corresponding neurons are indicated; therefore, the present embodiment records the offset relationship between the first output parameter value sequence of each neuron and the corresponding first loss function value sequence as the importance degree of each neuron in the first network.
Because the repeated training is performed on the same target training set, the learning requirement of the target training set of the next round can be predicted by utilizing the parameter adjustment gradient of the target training set during the previous round of training. Therefore, according to the importance degree of each neuron in the first network and the first parameter of each neuron corresponding to each sample silencer feature vector, the gradient is adjusted, and the adjustment coefficient of the first reference learning rate corresponding to each sample silencer feature vector is obtained; calculating an adjustment coefficient of a first reference learning rate corresponding to each sample muffler feature vector according to the following formula:
wherein F is an adjustment coefficient of a first reference learning rate corresponding to a certain sample muffler feature vector in the target training set, M is the number of neurons in a first network, Y m G is the importance of the mth neuron in the first network m Adjusting gradient for a first parameter of an mth neuron corresponding to the silencer feature vector in a target training set, wherein tau is a super parameter; y is Y m *G m The larger indicates the greater the learning requirement of the mth neuron on the muffler feature vector.
In this embodiment, multiplying an adjustment coefficient of a first reference learning rate corresponding to each sample muffler feature vector in the target training set by the corresponding first reference learning rate to obtain a first target learning rate of each sample muffler feature vector in the target training set; training the first network based on the first target learning rate and the target training data to obtain a second loss function value sequence corresponding to the second network and the target training set; then obtaining the stability degree of the second loss function value sequence, judging whether the stability degree of the second loss function value sequence is smaller than a preset stability degree threshold value, if so, calculating a second reference learning rate corresponding to each sample muffler characteristic vector in the target training set and an adjustment coefficient of the corresponding second reference learning rate; obtaining a second target learning rate corresponding to each sample muffler feature vector in the target training set according to the second reference learning rate corresponding to each sample muffler feature vector in the target training set and the adjustment coefficient of the corresponding second reference learning rate; training the second network based on a second target learning rate and target training data corresponding to the feature vectors of the silencer of each sample to obtain a third network; and analogically, stopping network training until the stability degree of the loss function value sequence is greater than or equal to a preset stability degree threshold value, and obtaining the target network.
In this embodiment, the method for obtaining the second loss function value sequence corresponding to the target training set is the same as the method for obtaining the first loss function value sequence corresponding to the target training set, so that detailed description is omitted. The method for obtaining the stability degree of the second loss function value sequence, the second reference learning rate corresponding to each sample muffler feature vector in the target training set and the adjustment coefficient of the corresponding second reference learning rate are the same as the method for obtaining the stability degree of the first loss function value sequence, the first reference learning rate corresponding to each sample muffler feature vector in the target training set and the adjustment coefficient of the corresponding first reference learning rate, so that the method is not described in detail; the method for obtaining the second target learning rate corresponding to the feature vector of each sample muffler in the target training set is the same as the method for obtaining the first target learning rate corresponding to the feature vector of each sample muffler in the target training set, and therefore will not be described in detail.
Step S003, obtaining data to be detected; and inputting the data to be detected into a target network to obtain the performance grade of the automobile air conditioner silencer corresponding to the data to be detected.
In this embodiment, data to be detected is obtained, where the data to be detected is a muffler feature vector corresponding to any automotive air conditioner muffler at any collection time, and the obtaining mode of the data to be detected is the same as that of each sample acoustic feature vector in the target training set; inputting the data to be detected into a target network to obtain the performance grade of the automobile air conditioner silencer corresponding to the data to be detected output by the network; the staff can evaluate the performance level of the silencer based on the performance level of the silencer of the automobile air conditioner corresponding to the obtained data to be detected.
The beneficial effects are that: in the embodiment, the target training set is used as a basis for obtaining a target network; taking the target network as a basis for obtaining the performance grade of the automobile air conditioner silencer; according to the embodiment, the network learning rate is adjusted based on the loss function value of the network in the multi-round training process and the network parameters during training, so that the network can be trained efficiently and precisely, and the performance of the automobile air conditioner silencer can be detected relatively reliably based on the target network.
The above embodiments are only for illustrating the technical solution of the present application, and are not limiting; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present application, and are intended to be included in the scope of the present application.
Claims (5)
1. An artificial intelligence-based air conditioner muffler performance detection method is characterized by comprising the following steps:
acquiring a target training set;
performing network training based on the fixed network learning rate and the target training set to obtain a first loss function value sequence corresponding to the first network and the target training set; judging whether the stability degree of the first loss function value sequence is smaller than a preset stability degree threshold value, if so, obtaining a first target learning rate of each data in a target training set according to the first loss function value sequence; training the first network based on the first target learning rate and the target training set to obtain a second network and a second loss function value sequence corresponding to the target training set; judging whether the stability degree of the second loss function value sequence is smaller than a preset stability degree threshold value or not, if so, obtaining a second target learning rate of each data in a target training set according to the second loss function value sequence; training the second network based on the second target learning rate and the target training set to obtain a third network and a third loss function value sequence corresponding to the target training set; and so on; stopping network training until the stability degree of the loss function value sequence is greater than or equal to a preset stability degree threshold value, and obtaining a target network;
acquiring data to be detected; inputting the data to be detected into a target network to obtain the performance grade of the automobile air conditioner silencer corresponding to the data to be detected;
the target training set comprises a plurality of sample muffler feature vectors;
performing network training based on the fixed network learning rate and the target training set to obtain a first loss function value sequence corresponding to the first network and the target training set; judging whether the stability degree of the first loss function value sequence is smaller than a preset stability degree threshold value, if so, obtaining a first target learning rate of each data in a target training set according to the first loss function value sequence, wherein the method comprises the following steps:
performing performance grade labeling on each sample muffler feature vector in the target training set;
training based on the characteristic vectors of the various sample silencers and the labeling performance grades of the characteristic vectors of the various sample silencers to obtain a first network;
calculating to obtain squares of differences between marking performance grades corresponding to the characteristic vectors of the sample mufflers and network prediction performance grades in the first network process, and obtaining first loss function values of the characteristic vectors of the sample mufflers;
constructing and obtaining a first loss function value sequence corresponding to the target training set according to the first loss function value;
calculating the standard deviation of the first loss function value sequence, and recording the inverse of the standard deviation of the first loss function value sequence as the stability of the first loss function value sequence;
and judging whether the stability degree of the first loss function value sequence is smaller than a preset stability degree threshold value, and if so, obtaining a first target learning rate of each sample muffler characteristic vector in a target training set according to the first loss function value sequence.
2. The method for detecting the performance of an air conditioner muffler based on artificial intelligence according to claim 1, wherein the method for obtaining a first target learning rate of each sample muffler feature vector in a target training set according to the first loss function value sequence comprises the following steps:
obtaining feature vectors of each sample muffler corresponding to each labeling performance grade in the target training set;
obtaining a first difference value of each sample muffler feature vector in the target training set according to the Euclidean distance between any two sample muffler feature vectors corresponding to each labeling performance grade;
obtaining a second difference value of each sample muffler feature vector in the target training set according to the average value of the sum of Euclidean distances between each sample muffler feature vector corresponding to each labeling performance level and each sample muffler feature vector corresponding to the rest labeling performance levels;
obtaining the performance grade distinguishing difficulty corresponding to each sample muffler feature vector according to the ratio of the first difference value of each sample muffler feature vector to the second difference value of the corresponding sample muffler feature vector in the target training set;
obtaining a first reference learning rate corresponding to a target training set according to the first loss function value of each sample muffler characteristic vector and the performance grade distinguishing difficulty of each sample muffler characteristic vector; obtaining a first reference learning rate of each sample muffler feature vector according to the first reference learning rate corresponding to the target training set;
obtaining a first output parameter value of each neuron corresponding to each sample silencer feature vector in the target training set and a first output parameter value sequence of each neuron corresponding to the target training set according to the process of obtaining the first network by the target training set;
obtaining a process of a first network and an inverse feedback network according to the target training set, and obtaining first parameter adjustment gradients of neurons corresponding to the characteristic vectors of the silencer of each sample in the target training set;
obtaining the importance degree of each neuron in a first network according to the absolute value of the partial correlation coefficient between the first output parameter value sequence and the corresponding first loss function value sequence;
according to the importance degree of each neuron and the first parameter of each neuron corresponding to each sample silencer feature vector, adjusting the gradient to obtain an adjustment coefficient of a first reference learning rate corresponding to each sample silencer feature vector;
and obtaining a first target learning rate of each sample muffler feature vector in the target training set according to the adjustment coefficient of the first reference learning rate corresponding to each sample muffler feature vector and the first reference learning rate corresponding to each sample muffler feature vector.
3. The artificial intelligence based air conditioning muffler performance testing method of claim 2, wherein the first variance value of each sample muffler feature vector in the target training set is calculated according to the following formula:
wherein,for a certain labeling performance level in the target training set>First difference value of characteristic vector of sample muffler, < ->For the first +.>Individual sample muffler characterizationQuantity and corresponding->Euclidean distance between individual sample muffler feature vectors,/->The number of sample muffler feature vectors in the performance level is noted for the target training set.
4. The artificial intelligence based air conditioning muffler performance detection method of claim 2, wherein the first reference learning rate corresponding to the target training set is calculated according to the following formula:
wherein,for a first reference learning rate corresponding to the target training set, < > for>Is super-parameter (herba Cinchi Oleracei)>For the number of sample muffler feature vectors in the target training set,/->For the target training set->Performance level discrimination difficulty of individual sample muffler feature vectors,/-for>For the target training set->A first loss function value of the sample muffler feature vector.
5. The method for detecting the performance of an air conditioner muffler based on artificial intelligence according to claim 2, wherein the method for calculating the adjustment coefficient of the first reference learning rate corresponding to each sample muffler feature vector according to the following formula comprises:
wherein,for the adjustment coefficient of the first reference learning rate corresponding to a certain sample muffler feature vector in the target training set,for the number of neurons in the first network, +.>Is the first->The degree of importance of the individual neurons,/-, and>for the first +.>First parameter of the individual neurons adjusts the gradient, < >>Is a super parameter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210432016.4A CN114838496B (en) | 2022-04-22 | 2022-04-22 | Air conditioner muffler performance detection method based on artificial intelligence |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210432016.4A CN114838496B (en) | 2022-04-22 | 2022-04-22 | Air conditioner muffler performance detection method based on artificial intelligence |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114838496A CN114838496A (en) | 2022-08-02 |
| CN114838496B true CN114838496B (en) | 2024-02-23 |
Family
ID=82565746
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210432016.4A Active CN114838496B (en) | 2022-04-22 | 2022-04-22 | Air conditioner muffler performance detection method based on artificial intelligence |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114838496B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115061050B (en) * | 2022-08-19 | 2022-11-01 | 南通博鎏机电科技有限公司 | Intelligent monitoring method for battery electric quantity |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007139232A (en) * | 2005-11-15 | 2007-06-07 | Panahome Corp | Ventilating structure |
| CN110285532A (en) * | 2019-07-04 | 2019-09-27 | 中国工商银行股份有限公司 | Method for controlling machine room air conditioner, apparatus and system based on artificial intelligence |
| KR20190119967A (en) * | 2018-04-13 | 2019-10-23 | 삼성전자주식회사 | Air conditioner and method for controlling air conditioner |
| JP2019220104A (en) * | 2018-06-22 | 2019-12-26 | 日本電信電話株式会社 | Device functioning as sensor node and data center, sensor network, communication method, and program |
| CN111076960A (en) * | 2019-11-22 | 2020-04-28 | 科士恩科技(上海)有限公司 | Voiceprint quality detection method based on artificial intelligence algorithm |
| CN111140986A (en) * | 2019-12-23 | 2020-05-12 | 珠海格力电器股份有限公司 | Operating state detection method and device of air conditioning system, storage medium and air conditioner |
| CN211177306U (en) * | 2019-11-12 | 2020-08-04 | 深圳孚沃德斯科技有限公司 | Intelligent energy-saving control system of air conditioner cooling system based on neural network |
| CN112097365A (en) * | 2020-07-10 | 2020-12-18 | 珠海派诺科技股份有限公司 | Air conditioner fault detection and identification method and device based on prediction and classification model |
| CN112254274A (en) * | 2020-10-21 | 2021-01-22 | 上海协格空调工程有限公司 | Air conditioner fault recognition system based on machine learning technology |
| CN113465139A (en) * | 2021-05-28 | 2021-10-01 | 山东英信计算机技术有限公司 | Refrigeration optimization method, system, storage medium and equipment |
| CN114119964A (en) * | 2021-11-29 | 2022-03-01 | 上海商汤临港智能科技有限公司 | Network training method and device, and target detection method and device |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107730000A (en) * | 2017-09-15 | 2018-02-23 | 珠海格力电器股份有限公司 | Coolant leakage detection method and device for air-conditioning |
| US20210023331A1 (en) * | 2019-07-22 | 2021-01-28 | Regents Of The University Of Minnesota | Computer architecture for identifying sleep stages |
| US20210158211A1 (en) * | 2019-11-22 | 2021-05-27 | Google Llc | Linear time algorithms for privacy preserving convex optimization |
-
2022
- 2022-04-22 CN CN202210432016.4A patent/CN114838496B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2007139232A (en) * | 2005-11-15 | 2007-06-07 | Panahome Corp | Ventilating structure |
| KR20190119967A (en) * | 2018-04-13 | 2019-10-23 | 삼성전자주식회사 | Air conditioner and method for controlling air conditioner |
| JP2019220104A (en) * | 2018-06-22 | 2019-12-26 | 日本電信電話株式会社 | Device functioning as sensor node and data center, sensor network, communication method, and program |
| CN110285532A (en) * | 2019-07-04 | 2019-09-27 | 中国工商银行股份有限公司 | Method for controlling machine room air conditioner, apparatus and system based on artificial intelligence |
| CN211177306U (en) * | 2019-11-12 | 2020-08-04 | 深圳孚沃德斯科技有限公司 | Intelligent energy-saving control system of air conditioner cooling system based on neural network |
| CN111076960A (en) * | 2019-11-22 | 2020-04-28 | 科士恩科技(上海)有限公司 | Voiceprint quality detection method based on artificial intelligence algorithm |
| CN111140986A (en) * | 2019-12-23 | 2020-05-12 | 珠海格力电器股份有限公司 | Operating state detection method and device of air conditioning system, storage medium and air conditioner |
| CN112097365A (en) * | 2020-07-10 | 2020-12-18 | 珠海派诺科技股份有限公司 | Air conditioner fault detection and identification method and device based on prediction and classification model |
| CN112254274A (en) * | 2020-10-21 | 2021-01-22 | 上海协格空调工程有限公司 | Air conditioner fault recognition system based on machine learning technology |
| CN113465139A (en) * | 2021-05-28 | 2021-10-01 | 山东英信计算机技术有限公司 | Refrigeration optimization method, system, storage medium and equipment |
| CN114119964A (en) * | 2021-11-29 | 2022-03-01 | 上海商汤临港智能科技有限公司 | Network training method and device, and target detection method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114838496A (en) | 2022-08-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110097755B (en) | Highway traffic flow state identification method based on deep neural network | |
| CN114838496B (en) | Air conditioner muffler performance detection method based on artificial intelligence | |
| CN104538041A (en) | Method and system for detecting abnormal sounds | |
| CN111504644B (en) | Bearing fault diagnosis method for adaptive anti-domain based on embedding distinguishability condition | |
| CN106649709B (en) | Vehicle stopping behavior pattern prediction and evaluation method based on data mining | |
| CN111024898B (en) | Vehicle exhaust concentration standard exceeding judging method based on Catboost model | |
| CN108182500A (en) | Ammunition Storage Reliability Forecasting Methodology based on accelerated life test | |
| CN110570013B (en) | Single-station online wave period data prediction diagnosis method | |
| CN111353373A (en) | Correlation alignment domain adaptive fault diagnosis method | |
| CN117173913B (en) | Traffic control method and system based on traffic flow analysis at different time periods | |
| CN115659782A (en) | Bridge damage identification method based on PGDNN under action of mobile vehicle load | |
| CN111967308A (en) | Online road surface unevenness identification method and system | |
| CN111062093A (en) | Automobile tire service life prediction method based on image processing and machine learning technology | |
| CN111027618A (en) | Automobile dynamic property and economic property expectation quantification method | |
| CN117390378B (en) | Intelligent management method and system for dual-carbon platform data | |
| CN114201632A (en) | Label noisy data set amplification method for multi-label target detection task | |
| KR20220078456A (en) | Method for air purifier filter life prediction based on ensemble supervised learning | |
| CN116609440B (en) | Intelligent acceptance management method and system for building engineering quality based on cloud edge cooperation | |
| CN110657355B (en) | Method for detecting leakage of thermal pipeline | |
| CN111476363A (en) | Stable learning method and device for distinguishing decorrelation of variables | |
| CN116910526A (en) | Model training method, device, communication equipment and readable storage medium | |
| CN115617882A (en) | Time sequence diagram data generation method and system with structural constraint based on GAN | |
| CN114660378A (en) | Multi-source detection parameter-based contact network comprehensive diagnosis method | |
| CN111012306B (en) | Sleep respiratory sound detection method and system based on double neural networks | |
| CN113159395A (en) | Deep learning-based sewage treatment plant water inflow prediction method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20240118 Address after: 1607, 16th Floor, 2nd Floor, Courtyard 18, Suzhou Street, Haidian District, Beijing, 100089 Applicant after: BEIJING BAICHEBAO TECHNOLOGY CO.,LTD. Address before: 226200 No.2, xiaohuaxiao Industrial Zone, Chengxi, Qidong City, Nantong City, Jiangsu Province Applicant before: JIANGSU FENGSHEN AIR CONDITIONING GROUP Co.,Ltd. |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |