TW564314B - CMAC_based fault diagnosis of air condition systems - Google Patents

Abstract

In this invention a CMAC_based fault diagnosis method for air condition system is presented. By using the known fault type mapping table as the training patterns to train the CMAC neural network, it can diagnose the possible fault type of large air condition system. Since the characteristic of association, generalization and the similar inputs activate similar memory addresses, this invention achieves at least the following merits: (1) high diagnosis accuracy is obtained. (2) High training and diagnosis speed. (3) High noise rejection ability. (4) Suit to multiple faults diagnosis and associate with the most similar fault type.

Description

564314 V. Description of the invention (i) [Field of the invention] The present invention is mainly related to the large-scale air-conditioning system, i-large-scale air-conditioning system by setting dozens of signals = diagnostic method. The state of the detection point judges the possible failure of the system, and advances the points to reduce the possible economic loss according to each signal ... When the signal detection is excluded, there are two or more failure judgments when corresponding codes are used. When a novel Fault diagnosis method can be correct :. The present invention is a kind of method, which is used as a basis for the maintenance of the staff and troubleshooting of the faults involved. [Background of the Invention] Failures of Internal Fans [Background of the Invention] For large-scale power equipment systems, such as airborne instruments, etc., there is a long history. ^ Power distribution system or electricity elimination to recover Let's talk about the operation of a spear and gray system to reduce the economic finger-thinking system as an example. The current microcomputer supervisory office I. The display of the warning is displayed by the two-tone display, and since the second control engineer 6 weeks, the system has both operation control,-automatic protection at the mouth, and system operating parameter measurement. Fu / huanli recording (such as ice-cold refrigerant return to low pressure, indoor and outdoor temperature, etc.) and other functions, it is still wonderful to rely on the purple code query during troubleshooting to find the old P ^ > staff! It can be judged based on the failure of the specific fault group or the early cause. As shown in Table 1, it is the same as the ice water machine. When ^ @ table, the detection signal corresponding to each coding bit is shown in Table 2. ; None: ;;: When, the maintenance personnel must comprehensively refer to various signal detections. Currently available: 2: Technical only can determine the possible cause of the failure. In large-scale systems, iik u. ^ ^ ^ There are many U, there are often more than dozens of ^ detection points, and the number of types is more than tens of millions (the signal combination of Table 1 Juma 24U). Due to the combination of the female G field and the corresponding fault code, the fault type is judged. Page 5 564314 V. Description of the invention (2) The difficulty is high, the dependence on expert experience is high, and it takes too much time and manpower. In order to quickly find out the cause of the system failure, discuss the techniques of fault diagnosis, such as the fault of the transformer, fault diagnosis in the literature, and fault diagnosis of the power system. ^ Withdrawal system methods, such as fault diagnosis of transformers, due to detection: analog signal detection is 'easy to directly use its method')) Air conditioners with more than dozens of binary signals at the measurement points are switched to inspection and training, training The time is too long, and the diagnosis is slow ...). = (3 The memory valley limit J = breaking technology, such as the establishment of fuzzy criteria must be seen in the second article to achieve a better diagnostic result. Generally, more sounds f seek, a Examination Moon Dagger

^ ^ t „t, 4 ^ 4 i (;;) f 7er) M ability will be lower, training must be all;" η two = obstacle type samples, easy to produce mutual / different reasons The value can be adjusted, the training time is long, and the best memory weight can be updated in the time of ': Ί pair ownership =. [Objective of the invention] The purpose of the invention is to introduce the cerebellum-like neural network CCMAC: Cerebel lar ModM 〇i隹 o odel Articulation Controller) fault diagnosis of air-conditioning system with ΓΜΑΓ I m detection signal. The use of 丨 will be A 4 / link, induction, local weight adjustment, similarity input signal multi-purpose Recalling the characteristics of the body address, we developed a set of digital detection methods that are suitable for digital detection: "": The system's fault detection method to achieve improved fault diagnosis. ^ Reduce training time, fast diagnosis, and reduce the

1 Page 6 564314 V. Description of the invention (3) __ Dependency and research goal of improving signal fault tolerance. ^ Used for multiple fault diagnosis of the system. [Inventive features] In order to achieve the purpose disclosed in the present invention, the feature is based on CMAC cerebellar neural network inflammation and its efficacy. The main technology and learning of the present invention. The similarity input signal via the network will be trained by the novel digital coding 2 structure for fault feature training to achieve a cerebellar-like model, which can effectively reduce the characteristics of similar memory when detecting a large number of detection signals, and can be divided into groups. The encoding method establishes the memory usage of the class [桓,]. First, the fault judgment table of the general air-conditioning chiller H = network architecture ', and secondly, the unified fault diagnosis. After the actual evaluation, the air-conditioning system or the newly added fault samples may be broken, if there is a misjudgment value. It is also worth checking I to update the best memory-right cerebellar model ^ to M: So the early diagnosis of P is converted to a classification problem, and the characteristics of CAMP and J-knowledge mode are used to classify & | classification functions, such as ί: Two human cerebellum carry out to identify who this person is. At 21, it can be easily judged to be with Γ7. If the individual is wearing a blindfold due to eye disease, or: To be able to tell who this person is. Even only seeing the eyes: Special: IT knows who this person is. The structure of the human brain is to-individual: ίίf cerebellar cells of a specific group 0, when seeing -individual only group? There are enough cerebellar cells to be compared (excited). Put on an eye mask; although the mask obstructs some features, the other features, and the rest of the features, stimulate the brain about this particular group

Page 7 564314 V. Description of Invention (4) A clear judgment can be made. After the judgment is completed, for example, eye masks, masks, etc. can be further converted into the basis for further judgment of this particular person in the future. CMAC is imitating τ; but it has the same characteristics and is very suitable for the cerebellum model, which shows the classification of the sects. And because of its cerebellar-like characteristics, Japanese scholars can better improve the ability of Ning, Yuan and Fen Fen (anti-noise). ί ^ Teaching Λ / When learning, the system only adjusts for specific groups of brain cells (excited memory addresses), such as = "_ Mo type or adaptive fuzzy logic V tune 4 ϊhibi Γ can understand The technical features of the present invention and its novel ν broken wooden structure and bismuth are illustrated as follows. [Introduction to Digital CMAC Neural Network] 姑姑 1 If imitate the learning structure of the human cerebellum, the bean is based on the received information. The input is inspired by the turn-in signal of the female daughter Shengzi Baikouji ^ ,,, and octave (specific group:… 4 differences to stimulate associative memory memory (brain cells) are more similar to fn 1). Large capacity / & almost the same as the structure of the human cerebellum, although its memory of a small part of t: "definite input signal, but only use the-address (a brain_ p MAC memory structure, each memory Body position ^ ^ ^ ^ ~ M ^ ^ ^ ^ ^ ^ can map out the oblique ^, and the weights in the body position are summed, that is, the value is equal to the ideal, and the output of the number l (Characteristics). The size of the difference between the output of this output and the one that is excited is divided evenly between this one " address adjustment (tunning), to complete training 564314 V. Invention Description (5) training. Therefore, when the same signal is re-input, because it will excite the same memory address, the sum of the memory When the weight value is obtained, the ideal output signal can be obtained. When the signal is accompanied by noise (the larger the noise, the lower the similarity with the original input signal), the original memory may only be partially re-activated. It is excited, so the sum of the excited memory output still retains some of the characteristics of the original output signal (depending on the similarity). Through the setting of threshold 7 ?, the distortion of the input signal is within a certain range Can still get the correct judgment. At the same time, based on the distortion of the ideal output signal of this distorted output signal, the weight of the excited note u can be further adjusted to further convert this distortion into the original input heart sign and again When the same distorted signal is input, it can be equal to the median value. The operating structure of the digital CMAC network can be graphed _ out of the space to the output signal, and the relationship between A and A is rotated into each group. In order to obtain one-shot = signal grouping, you can sum up the weights of the transmitted memory address; you can use the 7-out signal to calculate the error between the excited signal of each group and the ideal output signal. Weight basis. As shown in Figure 1, it is used to adjust the number of excited elements to 20, every five bits are divided into a group, then i assume that the bit code of the input signal excites the weight of four memory addresses, , = Input space Xl can be programmed A '(W' represents the ith address of the i-th group) „, < 乂, you can get-group output vector. This output vector :: the weight of four addresses: The error value is evenly distributed to these four addresses to compare the output data with the training data. Therefore, inputting ^ 1 again can complete a correct output. Enter all the data in sequence, and diagnose I to complete the error value based on the rotation error. 564314 V. Description of the invention (6) One training of data, set an evaluation criterion, up to b yuan training (the number of training indicates all data Repeating the strict index means that the CMAC system distributes a certain signal characteristic to the number of times that μ is entered). That is, if all are excited, you can determine this particular action, position, and these position settings are excited, it can only indicate that it is similar to the signal, and j is determined by the proportion of only a few bits. If the setting is high or low, the degree of similarity will be determined by excitement. If a certain bit is used to measure the signal X! When it is used to identify the signal as the 0th bit of the 1st group, then the generated error (such as the graph <, <, w %, In which only the address of the first group is changed, and at least 75% of its output characteristics are compared. The suitable user will be described later). Therefore, the increase in the number of memory locations used to store signal characteristics " ^% The effect of the number error of the cluster on the output signal will be reduced. Also g 9 plus, a single bit message outputs a high percentage of correct features, fault diagnosis ^ noise, at that time, it can still be high. For a trained CMAC network, the required fault-tolerant capabilities will be added to the final calculation. When the speed-up diagnosis performed by the user is only the group coding and the command input, similar reflections are taken & such as the cerebellum Receive similar speeds much faster than other intelligent diagnostic systems. In the study and diagnosis of the [Small and Small Mode Air Conditioning System Fault Diagnosis Architecture] As shown in Table 1, the fault judgments are as follows, and there are 44 kinds of early types. Check (see Table 2 for details) ). If the tested data are all fish = 4.: = Lian-bei materials are consistent, then the type of failure can be clearly identified and ": P is ruled out early. But 40 binary detection signals indicate that the number of combinations is 24. Therefore, the detected signal coding may not have a corresponding fault type. And the use of this comparison table can only perform a single fault diagnosis, not 564314 V. The ability to explain (7). The present invention is directed to a unified fault diagnosis method to solve the problem of cerebellar mode air conditioning systems such as Yushe's, and, and the lack of multiple faults. The broken and no fault code pairs are shown in Figure 2 as the input of the present invention. The signal is a 40 binary neural network model. This model number is divided into 8 groups (5 bit bluffs. The present invention takes the 40 letter bodies as a total of two parallel and parallel groups, saying: each- The memory 对应 corresponding to the group. That is, when the number of addresses of the rounder f is 32 layers of memory for training, when the second 锸 sample is input, only the first and second layers of memory of all groups are trained. The obstacle sample time is only for the entry of each group, the early characteristics of the same radical P of each layer. When the weights of the data to be diagnosed are added, the memory weights can be added individually: '== 2: send bit:'). The output value of （p of the output node (a total of 44 outputs are the same as the code of exhaustion, which means that the fault is V. It can be yes / no, a failure probability value, and the maintenance person g lies in the output representative of the δ > fault system. (For example, the closer the output is to the probability value, the more detailed the training, diagnosis, and efficacy of the fault = the type of failure.) This- Fault structure [general training of CMAC network] address state ::; order = two, grouping and stimulating for ㈤c; practice = :::::::: force; it should always be expressed in b. This = _ 1ΗΊ Page 11 564314 V. Description of the invention (8) Table 1 The input binary data of the first failure sample is

00101001001100110001OOOOOOOOOOOOOQQQOO With 5 bits as a group, there are 8 excitation addresses sequentially encoded from the least significant bit 1SB to the most significant bit MSB, in the order of al = a2 = a3 = a4 = 00000B = 0, a5 = 10001B = l7, a6 = 11001B = 25, a7 == 00100B = 4, a8 = 001〇1B = 5. Assuming the initial weights of all the memories are 0, the memory weights W% excited by the first sample are summed up , W ° 2, w ° 3, W ° 4, W175, w256, W47, w58 are 0. The output of CMAC can be expressed as

y = Σ w ', i = l, · · ·, a * A *: Number of memory addresses to be excited (1) Assume that output 1 represents a certain fault, so there is a clear output target (teacher) when discriminating with CMAC. The adopted learning rule is supervised learning, and the adjustment of each weight can directly use the steepest descent method, as shown in the following formula

Wail (new) -Wail (〇ld) + β (yd-y) / A *, (2)

Where W iUew} is the new weight after the adjustment of the stimulus memory, ㈤d) is the old weight before the adjustment of the stimulus memory, ai is the address of the stimulus memory, not the learning gain (0 < / 5 $ 1) , Yd is the target value (set by 丨 in the present invention), and y is the actual output value. It is worth noting that, if there is only one set of samples for each fault type, Lu Ke can be set directly. For more than one sample, the value of / 3 is usually slightly less than 1. The amount of memory required for each layer is related to the number of bits (m) in the cluster. The total amount of memory used is related to the number of groups, 5-in binary, number of bits (Π), and type of failure (f). Taking the structure of 刖 in Figure 2, A * = 8, m = 5, n = 40, f = 44, then the total memory address is

564314 V. Description of the invention (9) 44 X 25, the total memory MtQtal can be expressed by the following general formula: Mttai = A * x fx 2m = ce i 1 (n / m) xf χ 2m (3) where ceil (x The function) finds the closest integer to infinity. 3) n / m is not necessarily divisible, cei 1 indicates the function of unconditional carry, that is, the number of insufficient bits in the highest bit will be automatically rounded up when performing grouping. [Description of Fault Tolerance] 10,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000, and " a2, a3 , a4, a5, a6, a7, a8) from (〇, 〇,, H17,25,4,5), where only the al 'feature is allocated and stored in 8 locations, and only the feature output is still at least 8 or more, If you increase the number of clusters, the characteristics of the fault, and the percentage of errors detected by adjacent bits that have the correct amount of output will increase. This pattern is identical. The human cerebellum can be classified in a wide range, otherwise it will not affect its judgment related to the amount of memory. The resolution, the diversity and relevance of the failure sample data depend on the designer. The designer can only propose a The diagnosis architecture proposed by the present invention has good fault tolerance. The reasons can be explained as follows. Taking the first set of fault samples as an example, if two bits generate an error-a correct value: 001010010011001 error value: 001010010011001, the encoded excitation address (a 1, 0, 17, 25, 4, 5) becomes (3 , 0 produces an error. And because an address of the fault generates an error, the fault has a higher tolerance for fault tolerance, and the CMAC cerebellum has the ability to make modest failure samples unless it interferes with the proper analysis of the system. Similar adjustments to the number of clusters, such as obstacle diagnosis, are based on the ability to ensure that the impact of the allocation of storage reserves will reduce the behavior and identification of individuals.

V. Description of the invention (ίο) Example description [Convergence of CMAC training] The convergence proof of CMAC teaching learning has been found in the premise of no collision (co 1 1 isi on: different letters #: To the same memory), its convergence can be guaranteed. This "two = / 屮 digital diagnosis architecture, because the input signal is digital two & \ out the same as the analog input signal, there is no problem of collision, so learning is admirable. [Learning effectiveness evaluation]

Suppose that the memory of layer l (i = 1, ..., 44) is excited and the address weights 蛐 and ^ are output as 1, which means chapter ⑷. The number of data items of the i-type failure pattern sample is h. Let the evaluation index (4) Let E < ε, ε be greater than once E < ε is established, then E = Σ (yi-1) 2, i = 1, ··, ni, the value of E can indicate whether the learning effect is good or not. The number can be used as an evaluation of the learning effect to stop training. 〔Online Learning Mode〕

The advantage of the human small continuous cropping model is that it learns from mistakes, with Lian = the same mistake. Based on the recognition of people, the previously familiar face 〒 put on the blindfold ′ due to eye diseases hinders the comparison of some features, so it can be misjudged. However, if we have experienced one-time wrong discrimination, this eye mask feature ^: can be further converted into the unique features of this familiar face. Therefore, through learning, human beings can make progress in a certain day because of the accumulation of experience. The # 断 结构 of the present invention also mimics human behavior patterns.

564314 V. Description of the invention (11) Learning during misdiagnosis.

After completing the training network architecture, it is assumed that there is a new diagnosis for troubleshooting. If the data to be diagnosed and the training samples are turned in, correct fault diagnosis cannot be produced. What is wrong is wrong, and training can be adjusted to modify the past memory weight. The second is to do online and offline, mainly to adjust the output value to less than or greater than the threshold value, such as VaiKnew) = W ^ i (〇id) + α (^ -yerr) / A% ^ V. Among them, yerr is the output value of error diagnosis, α value is slightly. (5) No fault is misjudged as a fault >;. For example, the original ^ wind double r was earlier, indicating yerr > π. By using (recalling the weight value, the re-acoustic medium M ^ can be adjusted to record the IJ so that the Dandu-like input signal is not misjudged. However, the original fault is diagnosed as no reason = tower 4% output value [diagnostic rule 〕 3 principle is the same. As mentioned earlier, the diagnosis and diagnosis process proposed by the present invention is shown in Figure 3. "Offline §ji | practice mode can be summarized as follows, 1 Step 1 establish a cerebellar-like model of the diagnosis system, No. 44 Layer memory and 44 output programs. Contains 40 binary input points and determines the number of clusters. Step 2 Step 3 Load training samples into the CMAC model and add them to get the output of each node. The output value and the ideal value (yd = 1) Weight adjustment. 'Perform clustering, address coding and value comparison. Compare' Use formula (2) to No 'then go to step 2. Yes, step 4 is all training data training completed next step. 564314 V. Description of the invention 02) Performance evaluation of less miscellaneous 5 learning results. If no, go to step 2. Right ^ e, the memory weights are archived. Steps 1 to 5 are the offline learning etiquette to determine the training time. The training time of this training data sample is completed instantaneously (less than; ^ The collected sample is only 44 groups Therefore, "online mode, ° After completing the training in offline mode, from the diagnosis of the fault, the diagnosis step = the shirt is broken, you can perform the air conditioning system. Step 6 Enter the last memory: Step 7 Enter diagnostic data. Step 8. Perform grouping and place value. Horse and sum up to get the output of each node, then go to step 10. If not, go to step 10, if there is a next amount of money a go to step 12. ◊ ° 疋, go to step 7. No, Bay 丨 J Step 11 uses the formula (5), — Step 1 〇 仃 § Online memory weight adjustment, go to Step 12 to store the latest memory weight, the diagnosis is over. Figure 2— Left ^ Once i ^ Diagnosis and learning mode: Lu 3 f line training mode, the right is the online execution process, the second — =% right dash indicates the diagnosis after the first system startup, just load Previous memory step 9 diagnosis Correct? If it is on page 16 564314 V. Explanation of the invention (13) The value 'perform the process on the right half. [Case analysis and online learning] The diagnosis network architecture proposed by the present invention can be generalized and plumbed. Perform fault diagnosis. Table 3 is used by CMAC. Second, the time required for the second training is less than 1 #, (Pentlumllt 1 parameter MATLAB programming). Figure 4 shows the memory distribution of the first group of memory I: after Figure. Enter the sample of the table—the value of the second sample into the network as shown in Table 4 (partial results), which confirms that the diagnosis is positive: 100%. From Table 4, it can be seen that the diagnosis proposed by the present invention has two heart = system. Represents-probability value, 丨. Represents the determination of failure_, each Y point turns out the possible values of other failures, for example, the fourth type in Table 4 is the first type, but it is also pointed out that it is the first " type of heart = 75 %. In order to test the practicability and fault-tolerance of the method proposed in the present invention, the test data are generated arbitrarily to generate non-training samples (Table 5-9, fe ', the data of the training samples are based on the data in the table— j 11 groups of non-training Or r the same group of bits to change the second diagnosis two = ::: ;; more: six for this U The data are used for diagnosis. There is no table output, so there is no confirmed output value. However, you can see what kind of failure is most likely. You can still get the highest probability if the two outputs The fault diagnosis of the 6th, 8th, and 2nd gray ^ There are 3 false detection points, all of which are distributed in different m. The probability of the fault type due to its wrong energy is reduced to 62 5% and 'so its most There are two types of most likely faults, depending on the system :: =: = Ming 564314 V. Description of the invention (14) Degree of j. The threshold value (or maximum value) can be set to enable the proposed architecture of the present invention. 'Applicable to the diagnosis of a variety of fault types. /, [Retraining with the added data] The final maximum value of the non-training data in Table 5 is determined. C It: = 13 These data can be regarded as new sample data. At this time, I 2 The training adjustment mechanism uses formula (2) for retraining = 2 error diagnosis uses formula (5) for weight correction. ί i1 Weight distribution map of a group of memory memory weights after retraining. = Two: In order to find that the newly added sample will have a weight on the memory, this is the number of samples: The more sample data of the J training, the more it can make the memory: if the cerebellar cell memory is distributed. With this-online, a, and two, the diagnosis system will increase with the increase of diagnostic sample data, the accuracy of the entire mouth diagnosis. Part of the diagnosis results of the retraining after retraining are entered in the table 7 and the data are entered as 0, and 9 is set, and the diagnosis MU threshold is set to 0. Partly greater than 丨. Through training performance, the two: 丨 = the probability value of seven, the turn-out value is as close as possible. The evaluation (the number of compilations is greater than 1) can make the famous doctor: =! Γ accumulate through countless experiences and finally become a feasibility that has been proven. However, due to the sample of the data; the limited samples collected are likely to cause diagnostic errors. In order to improve the diagnostic function of this system 564314 V. Description of the Invention (15) ___, it depends on the diagnosis experience. Is it sufficient to apply the cerebellar model-like neural network to the accumulated sample, and then it can show its excellent function. ^ Do n’t, cherish 'when the training data is sufficient', the diagnostic output of this diagnostic system is the type of the data input obstacle type of non-training samples, to facilitate the maintenance of the fault. The test data is referenced by round. For example, Table 5 shows the efficacy disclosed by the present invention. The result data indeed proves that in summary of the above-mentioned publications, compared with the technical assistance and high diagnostic requirements, this case is specifically for the scope of the present invention and covers the nature of the present invention. And the proposed or approved patented air-conditioning system 'has a variety of fault-breaking accuracy and other characteristics to file a patent claim in accordance with the law' to ensure the preferred embodiment of the application and the empty diagnosis provided by the patent application of the spirit of the present invention In the method announcement, the capability of fault diagnosis is obviously also requested. I implore people's rights and interests. I haven't seen the referral department because of this technology, I'd like to provide you with a review. However, the technology can be used to change the effect. It is deeper than any existing case and it is also new. The case does not need to improve the system's pace. The investigating committee can disclose the limitation of the issue or repair. The specific patent of the expert's technical error was given earlier, and the right to clarify is still contained. Page 19 564314 Schematic description [Schematic description] Table 1 Air conditioning system fault diagnosis data sample table 2 Air conditioning fault diagnosis sample signal detection source table 3 CMAC network parameter table IV. Diagnostic output of training sample data (partial results) Table 5: Non-training data samples (including error detection bits ... bold type) Table 6 Table 5. Diagnostic output of non-training sample data (partial results) Table Table 7 Diagnostic output after retraining (partial results) Table 8 Table 5 Diagnostic output after retraining (partial results) Figure 1 Digital cerebellar mode network diagram Figure 2 CMAC cerebellar mode air conditioning system fault diagnosis model Figure 3. CMAC-based air-conditioning system fault diagnosis flowchart. Figure 4. Memory weight distribution of the first group after initial training. Figure 5. Memory weight distribution of the first group after retraining on the line. [Element symbol description] A * Association (excitation) ） Number of memory m Number of bits in clustering π Total number of bits in CMAC neural network input signal f Number of failure types β Training Gain a retraining learning gain yd desired output value

Page 564314

Page 21

Claims (1)

564314 Six network broken section issued a mark recall-one entry number and note 2. Its measurement 3. Loss of patent application scope of patent application scope of an air conditioning system failure of ancient and earth. ^ Road (CMAC) is the diagnosis year: It is based on the cerebellum-like model nerves, including the structure of the cerebellar model neural network to diagnose the early failure of the air conditioning system, including 40 binary works: two 4 layers of t group 3 memory layers and 44 rotations. Section :: ... 1 input is to group the input numbers in groups to generate the address weights to = letter: total per-layer memory mapping 2 = the training flow of network memory weights, according to the output signal and Project weight # & 丄 "! Errors are evenly allocated to the excitation address of the mapping for training adjustments based on questionable values. The system's fault diagnosis guidelines are based on the input of 40 binary signal detection signals to the training completion. The cerebellar neural network outputs a set of output signals to determine possible fault patterns. The online neural network weight adjustment mechanism can retrain the weights of misjudged data to update the best at any time. Memory weight. According to the air-conditioning system fault diagnosis method described in the first item of the scope of patent application, the 40 input signal nodes of the middle-class cerebellar model neural network are binary check numbers, used to indicate whether the detection point signals are normal or not; 44 layers The memory is used to memorize 44 types of fault features of the air-conditioning system; the value of 44 output nodes represents the possible diagnosis of this type of fault. According to the air-conditioning system fault diagnosis method described in the second item of the scope of the patent application, the out-node value is the memory right of the memory address of each group of 4 to 4 layers that is excited. Page 22 564314 VI. Patent Application Values are summed up individually to obtain the output value of each layer of memory. t d: The air-conditioning system fault diagnosis method described in item 3, wherein the closer the value of each output point is, the higher the probability that the fault type is changed and memorized. w Think 5 · As the percentage of Jg commanded in the scope of patent application 1 ▲ j. ψ ^ ^ ^ ^ ^, the method of industrial fault diagnosis,: = = The number of ΐ can be determined by the number of input signal groups. , The number of bits into n is n, and the number of m bits is U / m) Take an integer with an unconditional round. Qiao Qun excitement. 6 groups of diagnostic methods for air conditioning systems as described in item 5, each-: combination, the value of which is the memory stimulated by the memory of each group 7. As for the scope of patent application! In the air conditioner described in the item, the weight training of the offline God I network is based on a method that maps out the output data: to the excited memory address to complete the 7-item sample resource. In the method for diagnosing air conditioning system faults, the eighth middle school adjusts the weight of the training data of the mother and reed + age VA *, where yd = 1 is the ideal output value, and the upper Φ 〇 < cold S 1 is the training gain, which is better / A value of 3 is 1. ^ 9. The number of training sessions of the air conditioner as described in item 8 of the scope of the patent application is based on the sample data, and the more samples there are, the more training sessions you need. 10. Failure of the air conditioning system as described in item 9 of the scope of patent application: disconnection method, Page 23 564314 VI. The termination of the number of training sessions in the scope of patent application can be evaluated by the sum of: and the size, the sum of the squared errors (E) of all trainees is less than a value, 1, the error caused by Dongbei The default value is 0.01 p, which indicates that the training is not a good task. The best one is as follows: • The determination of the fault diagnosis criteria is based on the application of the patent application. The obstacle diagnosis method sets a threshold value of π. Depending on whether the output value of π is greater than the pre-five requirement, the general threshold value is 0.8; for air <, the safety of the system's operation. / 、 Judgment of intermediate fault diagnosis criteria, secondary system fault diagnosis method, and judgment of fault type. The maximum output value of the neutral point is subject to empty grinding as described in item i of the patent application scope: the weight of the neural network on the midline The system adjusted the fault diagnosis method of the system, and adjusted the memory weight value. The dragon system was used for fault diagnosis, that is, the memory weight adjustment Awai, and the rotation value was adjusted to be smaller than the threshold α (, / P # i (new) = Wai. At 1 err, Yerr in Eight is wrong Misdiagnosis 11) Mountain, preferably 1.1. Output value, α value is slightly larger Page 24