AI intelligent diagnosis method and system based on intelligent system

Abstract

The invention provides an AI intelligent diagnosis method and system based on an intelligent system, and the method comprises the steps: collecting a building electromechanical equipment operation data flow containing multi-source heterogeneous data, such as a vibration spectrum, a thermal imaging characteristic and an energy consumption fluctuation parameter, through an edge calculation node, and carrying out the state time sequence coding to generate a coding sequence; and inputting the sequence into a fault prediction neural network to obtain a multi-dimensional fault feature distribution matrix. And then obtaining a standard diagnosis report of the expert knowledge base and an emergency diagnosis report of the field operation and maintenance record. And then a composite training objective function is constructed based on the result and equipment operation characteristic hidden space mapping, a fault prediction neural network parameter space is optimized through a back propagation algorithm, and a target fault prediction neural network is generated. And finally, fault prediction and diagnosis are carried out on the operation data flow of the building electromechanical equipment by using the network, so that comprehensive, accurate and intelligent fault diagnosis is realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more specifically, to an AI intelligent diagnosis method and system based on an intelligent system. Background Art

[0002] In the field of operation and maintenance of building electromechanical equipment, ensuring the stable operation of equipment is crucial for guaranteeing the normal use of buildings. Traditional fault diagnosis methods for building electromechanical equipment have many limitations and are difficult to meet the requirements of the increasingly complex electromechanical systems in modern buildings.

[0003] In the early stage, manual regular inspections were mainly relied on to detect equipment failures. Maintenance personnel inspected the equipment based on experience and simple tools, but this method was inefficient, difficult to monitor the equipment operation status in real time, and not easy to detect potential early failures. Often, the failures were not discovered until they had developed to a certain extent and had a significant impact on the equipment operation, which might lead to problems such as increased equipment damage, increased maintenance costs, and affected building-related functions.

[0004] With the development of technology, some sensor-based monitoring methods began to emerge. These methods usually only use a single type of sensor to collect data, such as only focusing on single-dimensional information such as vibration signals or temperature changes, and cannot comprehensively reflect the complex operation status of the equipment. Due to the diverse and interrelated fault modes of building electromechanical equipment, monitoring with a single data source is difficult to accurately diagnose the cause and location of faults, and is prone to misdiagnosis or missed diagnosis.

[0005] Later, some intelligent diagnosis technologies introduced machine learning algorithms. However, most of these algorithms directly perform simple classification or regression analysis after collecting a large amount of data, without fully considering the time-series characteristics of equipment operation data and the internal relationship between different types of data. For example, simply splicing the collected various data and inputting it into the model without effectively processing the data to mine the deep information therein, resulting in limited ability of the model to identify complex fault modes and low diagnostic accuracy.

[0006] At the same time, existing fault diagnosis methods have deficiencies in knowledge utilization. Either only rely on historical data to train the model, lacking the professional knowledge guidance of domain experts, making the model unable to give effective diagnostic results when facing some complex and rare faults; or simply rely on expert experience to formulate diagnostic rules, which are difficult to adapt to the continuously changing working conditions and newly emerging fault types during the equipment operation process, lacking flexibility and adaptability. Moreover, in the actual operation and maintenance process, the valuable real-time information and temporary disposal experience contained in the on-site operation and maintenance records have not been fully utilized, and these information cannot be combined with traditional diagnostic knowledge to optimize the diagnostic process. Summary of the Invention

[0007] In view of the problems mentioned above, in combination with the first aspect of the present invention, embodiments of the present invention provide an AI intelligent diagnosis method based on an intelligent system, and the method includes:

[0008] Collect the operation data stream of building mechanical and electrical equipment through an edge computing node, and perform state time series encoding on the operation data stream to generate an equipment state time series encoding sequence, where the operation data stream includes vibration spectra, thermal imaging features, and energy consumption fluctuation parameters collected by multi-source heterogeneous sensors;

[0009] Input the equipment state time series encoding sequence into a fault prediction neural network, and output a multi-dimensional fault feature distribution matrix, where the multi-dimensional fault feature distribution matrix includes the probability weights of equipment potential fault modes on preset diagnosis dimensions and the abnormal confidence levels of associated components;

[0010] Obtain a standard diagnosis report generated by an expert knowledge base and an emergency diagnosis report generated by on-site operation and maintenance records, where the standard diagnosis report includes accurate fault location information and composite repair strategies, and the emergency diagnosis report includes temporary disposal plans and unvalidated abnormal speculations;

[0011] Based on the multi-dimensional fault feature distribution matrix, the verification diagnosis unit of the standard diagnosis report, the unvalidated diagnosis unit of the emergency diagnosis report, and the mapping of the equipment operation feature hidden space, construct a composite training objective function;

[0012] Optimize the parameter space of the fault prediction neural network through the backpropagation algorithm. After the composite training objective function reaches the convergence threshold, generate a corresponding target fault prediction neural network, and perform fault prediction diagnosis on the operation data stream of any input building mechanical and electrical equipment based on the target fault prediction neural network.

[0013] In another aspect, embodiments of the present invention further provide an AI intelligent diagnosis system based on an intelligent system, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or codes, and the processor is used to execute the programs, instructions, or codes in the machine-readable storage medium to implement the above method.

[0014] Based on the above aspects, embodiments of the present application collect operation data streams such as vibration spectra, thermal imaging features, and energy consumption fluctuation parameters of building mechanical and electrical equipment obtained by multi-source heterogeneous sensors, and generate an equipment state time series encoding sequence through state time series encoding. Compared with traditional diagnosis methods using a single data source or simply splicing data, it can capture the complex features of the equipment operation state more comprehensively and accurately, organically integrate the key information contained in different types of data, lay a solid foundation for subsequent accurate fault diagnosis, and greatly improve the ability to identify equipment potential fault modes.

[0015] By inputting the device status time series coding sequence into the fault prediction neural network, a multi-dimensional fault feature distribution matrix containing the probability weights of the potential fault modes of the device on the preset diagnosis dimensions and the abnormal confidence degrees of the associated components is output, breaking through the limitation of traditional diagnosis methods that only give simple fault classifications or rough fault locations, presenting fault features in a more refined and quantitative manner, providing rich and valuable decision-making basis for diagnosticians, helping to more accurately judge the fault types and influence ranges, and improving the reliability and accuracy of fault diagnosis.

[0016] Obtain the standard diagnosis report generated by the expert knowledge base and the emergency diagnosis report generated by the on-site operation and maintenance records, and based on this, construct a composite training objective function with the multi-dimensional fault feature distribution matrix and the hidden space mapping of the device operation features, thereby combining the profound knowledge of domain experts and the actual on-site operation and maintenance experience, and overcoming the deficiencies of simply relying on historical data or a single diagnosis knowledge source. The accurate fault location information and composite repair strategy in the standard diagnosis report provide accurate diagnosis references for the model, and the temporary disposal plan and un-verified abnormal speculation in the emergency diagnosis report introduce the flexibility and real-time factors in the actual scenario, enabling the model to optimize training based on a comprehensive variety of information, and thus generating a target fault prediction neural network that better meets the actual needs and has greater practicality and adaptability.

[0017] Adopt the backpropagation algorithm to optimize the parameter space of the fault prediction neural network, and generate the target fault prediction neural network after the composite training objective function reaches the convergence threshold. This enables the fault prediction neural network to automatically adjust parameters during the training process to balance the influence of various aspects of information, not only improving the model's fitting ability for known data, but also enhancing its generalization ability for any input building mechanical and electrical equipment operation data stream. Compared with traditional diagnosis methods with fixed model structures or single optimization objectives, the target fault prediction neural network generated by this method can better adapt to the fault prediction and diagnosis of different working conditions and different types of building mechanical and electrical equipment, significantly improving the generality and applicability of the diagnosis method. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic flowchart of the execution process of the AI intelligent diagnosis method based on an intelligent system provided by an embodiment of the present invention.

[0019] Figure 2 It is a schematic diagram of the hardware architecture of the AI intelligent diagnosis system based on an intelligent system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] The present invention will be specifically described below in conjunction with the accompanying drawings of the specification. Figure 1It is a schematic flowchart of an AI intelligent diagnosis method based on an intelligent system provided by an embodiment of the present invention. The AI intelligent diagnosis method based on the intelligent system will be introduced in detail below.

[0021] Step S110: Collect the operation data stream of building mechanical and electrical equipment through an edge computing node, and perform state time series encoding on the operation data stream to generate an equipment state time series encoding sequence. The operation data stream includes vibration spectra, thermal imaging features, and energy consumption fluctuation parameters collected by multi-source heterogeneous sensors.

[0022] In this embodiment, in a large commercial building, there are numerous mechanical and electrical equipment, such as air conditioning systems, elevator systems, ventilation systems, etc. Edge computing nodes are deployed near these mechanical and electrical equipment. For the air conditioning system, multi-source heterogeneous sensors start to work. Among them, vibration sensors collect the vibration spectra of air conditioning compressors, which record the vibration conditions of the compressors during operation at a certain frequency. For example, 100 vibration data are collected per second, and these data reflect the motion states of the internal mechanical components of the compressors. Thermal imaging sensors collect thermal imaging features for the condensers and evaporators of the air conditioners, which can capture the temperature distribution on the surface of the equipment. For example, the temperature of some areas on the surface of the evaporator is relatively low, and if there are heat dissipation problems, local high-temperature areas may appear. Energy consumption sensors are responsible for collecting the energy consumption fluctuation parameters of the air conditioning system. Since the energy consumption of the air conditioner varies significantly under different operating modes (such as cooling, heating, dehumidification, etc.), the sensors record the energy consumption data at a certain time interval (such as every 5 minutes).

[0023] The edge computing node receives the data from these different sensors to form an operation data stream. Then, state time series encoding is performed on this operation data stream. Taking the vibration spectrum data of the air conditioning compressor as an example, the continuously collected vibration frequency values are encoded in chronological order. Suppose the vibration frequency is 50Hz at a certain moment and 51Hz at the next moment. Then, when encoding, these two time points and the corresponding frequency values will be converted into a part of the equipment state time series encoding sequence according to a specific encoding rule. Similar methods are also used for encoding thermal imaging features and energy consumption fluctuation parameters. Finally, a complete equipment state time series encoding sequence is generated, which completely records the operation state information of the building mechanical and electrical equipment over a period of time.

[0024] Step S120: Input the equipment state time series encoding sequence into a fault prediction neural network, and output a multi-dimensional fault feature distribution matrix. The multi-dimensional fault feature distribution matrix includes the probability weights of equipment potential fault modes in preset diagnosis dimensions and the abnormal confidence degrees of associated components.

[0025] Taking the elevator system in a building as an example, the previously generated device status time series encoding sequence is input into the fault prediction neural network.

[0026] First, the device status time series encoding sequence is subjected to modal separation through a multi-channel time series decomposition module. For the elevator system, the multi-source heterogeneous sensor data it contains is also decomposed. For example, the vibration spectrum subsequence collected by the vibration sensor at the bottom of the car, which contains the vibration state encoding vector of the car during the elevator operation, and the timestamps are precisely aligned. At the same time, the thermal imaging feature subsequence of key components such as the motor collected by the thermal imaging sensor is also separated, where each encoding vector reflects the temperature distribution correlation feature of the component surface at the corresponding moment. There is also the energy consumption fluctuation parameter subsequence, which reflects the energy consumption changes of the elevator at different operation stages (such as ascending, descending, docking, etc.).

[0027] The vibration spectrum subsequence is input into the first time series feature encoder. The vibration data of the elevator car extracts local vibration waveform features through a time series convolutional network. For example, when there is slight wear on the car guide rail, specific frequency changes will be reflected in the local vibration waveform. At the same time, a bidirectional long short-term memory network is used to capture the long-period dependence relationship of the vibration mode. For example, when the elevator operates at different times of the day, its vibration mode may have periodic changes, which are all converted into vibration feature vectors.

[0028] The thermal imaging feature subsequence is input into the second time series feature encoder. For the elevator motor, the spatial correlation feature of the thermal distribution on the motor surface is extracted through a spatial attention mechanism. For example, if there is a blockage in the heat sink part of the motor, it will cause local overheating, and this spatial correlation feature will be captured. Then, a gated recurrent unit is combined to capture the dynamic propagation path of the temperature change. When a fault occurs inside the motor and the heat cannot be dissipated normally, the temperature will gradually rise, and this dynamic propagation process of the temperature change is converted into a thermal imaging feature vector.

[0029] The energy consumption fluctuation parameter subsequence is segmented by a multi-scale sliding window. For the elevator system, the short-term energy consumption trend segment may reflect the energy consumption changes during one docking and start-up process of the elevator, while the long-term energy consumption baseline segment reflects the average energy consumption level of the elevator over a longer period (such as a day). They are respectively input into the parallel branches of the third time series feature encoder. The short-term energy consumption mutation features are extracted through dilated convolutional kernels. For example, when there is a fault in the braking system of the elevator, there may be a mutation in the short-term energy consumption. The long-term energy consumption baseline features are fused through a self-attention mechanism to generate an energy consumption feature vector.

[0030] Construct a cross-modal feature fusion module to perform tensor concatenation on the vibration feature vector, thermal imaging feature vector, and energy consumption feature vector. For example, in an elevator system, the features of the vibration of the car, the temperature of the motor, and the overall energy consumption are interrelated. Calculate the dynamic correlation weights between them through the multi-head attention mechanism to generate a cross-modal correlation matrix, and then output the fused device state feature tensor through feature channel weighted pooling.

[0031] Input the device state feature tensor into the spatio-temporal feature aggregation module, combined with the elevator component topology structure diagram data. For example, there are physical connection relationships between components such as the car, the traction machine, and the guide rail. Aggregate the state propagation features of adjacent components through the graph convolutional network. When the traction machine fails, it may affect the operating state of the car through the connected components, and this state propagation feature will be aggregated. At the same time, adopt a timestamp alignment strategy to fuse the difference gradients between the historical state features and the real-time state features. For example, compare the operating state of the elevator in the past week with the current real-time state to generate a spatio-temporal aggregation feature matrix.

[0032] Construct a dynamic weight allocation layer based on the spatio-temporal aggregation feature matrix. For the elevator system, the contribution degrees of different feature dimensions in the preset fault mode discrimination space are different. For example, the vibration feature of the car has a greater contribution degree to judging the fault of the car guide rail. Calculate the contribution degrees of each feature dimension through a learnable parameter matrix to generate a feature weight vector, and use a soft threshold function to adaptively filter the features with low contribution degrees, and output the optimized high-dimensional fault feature mapping.

[0033] Input the high-dimensional fault feature mapping into the abnormal mode detection module. Assume that the state reference distribution during the normal operation of the elevator is pre-learned through a deep autoencoder. When an abnormality occurs, such as abnormal shaking of the car, calculate the abnormal area in the reconstruction error matrix that exceeds the preset threshold to generate a potential abnormal mode vector. And perform time window sliding matching on the abnormal mode in combination with the elevator operation context information. For example, considering the operating characteristics of the elevator when stopping at different floors, output a candidate fault mode set, which may include candidate fault modes such as car guide rail wear and traction machine failure.

[0034] Perform similarity measurement on each candidate fault mode in the candidate fault mode set and the standard fault feature map in the expert knowledge base. For example, compare the candidate fault mode of car guide rail wear with the features of guide rail wear in the standard fault feature map to generate an initial matching probability. Adjust the initial matching probability according to the cumulative occurrence frequency of the same type of fault mode in the elevator operation historical data. For example, if the frequency of guide rail wear faults is relatively high in history, then appropriately increase the matching probability of the current candidate fault mode to generate a dynamically weighted fault mode probability distribution.

[0035] Construct a belief propagation network based on the physical connection relationship and fault propagation path among elevator components. For example, the wear of the car guide rail may affect the running smoothness of the car, and further may affect the working state of the safety device. Perform belief diffusion calculation on the probability distribution of the fault mode, and generate a belief propagation matrix containing the fault influence factors among components by iteratively updating the abnormal belief weights of associated components.

[0036] Fuse the belief propagation matrix with the dynamically weighted probability distribution of the fault mode tensorially, and generate a multi-dimensional fault feature distribution matrix through the normalized exponential function. In this matrix, the row dimension represents the probability weight distribution on the preset diagnosis dimension, such as the probability weights of different types of faults (such as mechanical faults, electrical faults, etc.), and the column dimension represents the abnormal belief gradient of associated components, such as the abnormal beliefs of components such as the car and the traction machine.

[0037] Step S130, obtain the standard diagnosis report generated by the expert knowledge base and the emergency diagnosis report generated by the on-site operation and maintenance records. The standard diagnosis report contains accurate fault location information and composite repair strategies, and the emergency diagnosis report contains temporary disposal plans and un-verified abnormal speculations.

[0038] Still taking the elevator system as an example, the expert knowledge base is accumulated through years of elevator maintenance experience and theoretical research. When an elevator fails, the expert knowledge base will generate a standard diagnosis report based on the previous fault mode identification and analysis. For example, if it is determined that there is a wear fault in the car guide rail, the accurate fault location information may indicate that a certain section of the car guide rail (such as between the 5th floor and the 10th floor) is worn, and the composite repair strategy may include a series of operations such as replacing the worn guide rail components and adjusting the balance of the car.

[0039] The on-site operation and maintenance records are the emergency diagnosis reports recorded by on-site maintenance personnel in emergency situations. For example, when the elevator suddenly stops running, after the on-site maintenance personnel's preliminary inspection, they find that the car has a slight shake. They make some temporary disposal plans according to experience, such as first stopping the elevator at the nearest floor, evacuating the passengers, and then recording the un-verified abnormal speculations in the emergency diagnosis report. It may be speculated that there is a foreign object stuck in the car guide rail or there is a problem with the car balance, but these speculations have not been further verified.

[0040] Step S140, construct a composite training objective function based on the multi-dimensional fault feature distribution matrix, the verified diagnosis unit of the standard diagnosis report, the un-verified diagnosis unit of the emergency diagnosis report, and the implicit space mapping of the equipment operation characteristics.

[0041] For the elevator system, first calculate the reliability loss value of each verified diagnosis unit in the standard diagnosis report and the error correction value of each un-verified diagnosis unit in the emergency diagnosis report.

[0042] In a standard diagnostic report, each verified diagnostic unit contains the basis for judging a specific fault. For example, for the verified diagnostic unit of car guide rail wear, the equipment state discrimination nodes include the wear degree and wear position of the guide rail, etc. Obtain the credibility distribution of this equipment state discrimination node in the corresponding fault mode (car guide rail wear) from the multi-dimensional fault feature distribution matrix. At the same time, based on the dynamic deviation degree between the equipment operation baseline parameters and the real-time monitoring parameters, construct the confidence attenuation function of each discrimination node in the diagnostic unit. For example, under normal circumstances, there is a baseline parameter for the wear speed of the car guide rail. If the actually monitored wear speed exceeds this baseline, then the confidence will be attenuated according to this deviation degree. Calculate the pedigree matching loss between the credibility distribution of each discrimination node in the verified diagnostic unit and the preset fault feature map marked by experts. If there are specific manifestations when the guide rail wears to a certain extent stipulated in the fault feature map, and the actually monitored situation deviates from it, a pedigree matching loss will occur.

[0043] For the unverified diagnostic unit in the emergency diagnostic report, for example, it was previously speculated that there was a foreign object stuck in the car guide rail or there was a problem with the car balance. Statistically analyze the misjudgment frequency of these abnormal speculation nodes in historical operation and maintenance cases to generate a dynamic error correction weight. If there is a high misjudgment rate in past similar speculations, then this weight will be relatively low.

[0044] Dynamically weight and fuse the mean value of the reliability loss of the verified diagnostic unit and the mean value of the error correction of the unverified diagnostic unit to generate an optimized target component for diagnostic accuracy.

[0045] Then extract the high-order representation vectors of the standard diagnostic report and the emergency diagnostic report in the equipment feature latent space. For the standard diagnostic report, convert it into an equipment health state reference vector through a deep feature extraction network. This vector contains component life cycle characteristics and compound fault association patterns. For example, the expected life of the elevator car guide rail, the association relationship between guide rail wear and other component failures, etc. Map the emergency diagnostic report into a real-time operation and maintenance response feature vector. This vector contains the abnormal propagation path and the emergency measure impact factor. For example, if there is a problem with the car guide rail, it may affect the running smoothness of the car. This is the abnormal propagation path, and the temporary disposal measures taken on site (such as stopping to evacuate passengers) are the emergency measure impact factors. Calculate the cosine similarity between the equipment health state reference vector and the real-time operation and maintenance response feature vector in the fault traceability space, and dynamically scale the cosine similarity based on the equipment operation environment parameters. For example, environmental parameters such as the building height and usage frequency of the elevator will affect the calculation of this cosine similarity. Optimize the feature encoder of the fault prediction neural network through an adversarial training mechanism to make the cosine similarity reach the preset stability threshold, thereby generating an optimized target component for feature alignment.

[0046] Finally, non-linearly couple the diagnostic accuracy optimization target component and the feature alignment optimization target component to generate a composite training objective function.

[0047] In step S150, optimize the parameter space of the fault prediction neural network through the backpropagation algorithm. After the composite training objective function reaches the convergence threshold, generate the corresponding target fault prediction neural network, and perform fault prediction and diagnosis on the operation data stream of any input building electromechanical equipment based on the target fault prediction neural network.

[0048] Taking the elevator system as an example, after constructing the composite training objective function, use the backpropagation algorithm to optimize the parameter space of the fault prediction neural network. In each iteration process, adjust the parameters such as the weights and biases of the neural network according to the value of the composite training objective function. For example, if the value of the composite training objective function is large, it indicates that there is a large deviation between the current neural network prediction result and the actual situation (such as the standard diagnostic report and the emergency diagnostic report, etc.). Then, propagate the error from the output layer to the input layer through the backpropagation algorithm and adjust the parameters of each intermediate layer.

[0049] As the number of iterations increases, the composite training objective function gradually approaches the convergence threshold. When the convergence threshold is reached, the corresponding target fault prediction neural network is generated.

[0050] At this time, for any input operation data stream of the elevator system, such as the newly collected data including the car vibration spectrum, motor thermal imaging features, and energy consumption fluctuation parameters, etc., the target fault prediction neural network can perform fault prediction and diagnosis. It will process the input operation data stream according to the previously trained mode and output a multi-dimensional fault feature distribution matrix including the probability weights of potential fault modes on the preset diagnostic dimension and the abnormal confidence degrees of associated components, so as to accurately judge whether the elevator has a fault and the possible types and locations of the fault, etc.

[0051] Based on the above steps, the embodiments of the present application collect operation data streams such as vibration spectra, thermal imaging features, and energy consumption fluctuation parameters of building electromechanical equipment obtained by multi-source heterogeneous sensors, and generate equipment state time series coding sequences through state time series coding. Compared with the traditional diagnostic methods using single data source or simply splicing data, it can capture the complex features of the equipment operation state more comprehensively and accurately, organically integrate the key information contained in different types of data, lay a solid foundation for subsequent accurate fault diagnosis, and greatly improve the recognition ability of potential fault modes of the equipment.

[0052] By inputting the device status time-series coding sequence into the fault prediction neural network, a multi-dimensional fault feature distribution matrix is output, which includes the probability weights of potential device fault modes on preset diagnosis dimensions and the abnormal confidence degrees of associated components. This breaks through the limitation of traditional diagnostic methods that only give simple fault classifications or rough fault locations, presents fault features in a more refined and quantitative way, provides rich and valuable decision-making basis for diagnosticians, helps to more accurately judge the fault types and influence scopes, and improves the reliability and accuracy of fault diagnosis.

[0053] Obtain the standard diagnostic report generated by the expert knowledge base and the emergency diagnostic report generated by on-site operation and maintenance records, and based on this, construct a composite training objective function with the multi-dimensional fault feature distribution matrix and the implicit space mapping of device operation features. Thus, it combines the profound knowledge of domain experts and on-site actual operation and maintenance experience, and overcomes the deficiency of simply relying on historical data or a single diagnostic knowledge source. The accurate fault location information and composite repair strategy in the standard diagnostic report provide accurate diagnostic references for the model, while the temporary disposal plan and un-verified abnormal speculation in the emergency diagnostic report introduce the flexibility and real-time factors in the actual scenario, enabling the model to optimize training based on a combination of multiple pieces of information, and thus generating a target fault prediction neural network that better meets the actual needs, is more practical and adaptable.

[0054] Use the backpropagation algorithm to optimize the parameter space of the fault prediction neural network, and generate the target fault prediction neural network after the composite training objective function reaches the convergence threshold. This enables the fault prediction neural network to automatically adjust parameters during the training process to balance the influence of various aspects of information, not only improving the model's fitting ability for known data, but also enhancing its generalization ability for the operation data stream of any input building electromechanical equipment. Compared with traditional diagnostic methods with fixed model structures or single optimization objectives, the target fault prediction neural network generated by this method can better adapt to the fault prediction and diagnosis of different working conditions and different types of building electromechanical equipment, significantly improving the generality and applicability of the diagnostic method.

[0055] For example, in a possible implementation manner, step S120 includes:

[0056] Step S121, perform modal separation on the device status time-series coding sequence through a multi-channel time-series decomposition module to generate vibration spectrum subsequences, thermal imaging feature subsequences, and energy consumption fluctuation parameter subsequences corresponding to multi-source heterogeneous sensor types. Each subsequence contains device status coding vectors with time-stamp alignment.

[0057] In this embodiment, taking the ventilation system in a building as an example, the ventilation system includes data collected by various sensors, and these data constitute a time series coding sequence of device states. When performing modal separation, subsequences corresponding to multi-source heterogeneous sensor types will be generated. For example, the vibration spectrum data collected by the vibration sensor installed on the ventilator is separated into a vibration spectrum subsequence, and each device state coding vector in this vibration spectrum subsequence carries an accurate timestamp, which completely records the vibration state information of the ventilator at different times. The thermal imaging feature data collected by the thermal imaging sensor for the motor and heat dissipation components of the ventilator is separated into a thermal imaging feature subsequence, and the device state coding vectors therein accurately reflect thermal imaging features such as the temperature distribution at the corresponding time. In addition, the ventilation system energy consumption fluctuation parameters collected by the energy consumption sensor are separated into an energy consumption fluctuation parameter subsequence, and its device state coding vectors are also arranged in chronological order, covering the energy consumption conditions at different time points.

[0058] Step S122, input the vibration spectrum subsequence into the first time series feature encoder, extract local vibration waveform features through a time series convolutional network, and use a bidirectional long short-term memory network to capture the long-term dependencies of vibration patterns, generating a vibration feature vector. At the same time, input the thermal imaging feature subsequence into the second time series feature encoder, extract the spatial correlation features of the thermal distribution through a spatial attention mechanism, and combine a gated recurrent unit to capture the dynamic propagation path of temperature changes, generating a thermal imaging feature vector.

[0059] During the operation of the ventilator, its vibration spectrum contains rich information. The time series convolutional network will process the vibration spectrum subsequence to extract local vibration waveform features. For example, if there are slight imbalances in the blades of the ventilator, it will be reflected as fluctuations with specific frequencies and amplitudes in the local vibration waveform. At the same time, the bidirectional long short-term memory network will capture the long-term dependencies of vibration patterns. The operating load of the ventilator is different at different times of the day, and this load change will lead to long-term changes in the vibration pattern. Through this network, this long-term dependency can be captured, and then a vibration feature vector is generated.

[0060] At the same time, input the thermal imaging feature subsequence into the second time series feature encoder. For the motor and heat dissipation components of the ventilator, the spatial attention mechanism will focus on the thermal distribution in the thermal imaging feature subsequence and extract the spatial correlation features of the thermal distribution. For example, if there is a local short circuit in the winding part of the motor, it may show a high temperature in a specific area on the thermal imaging, and the spatial correlation features between this high-temperature area and the surrounding areas will be extracted. Then, the gated recurrent unit will capture the dynamic propagation path of temperature changes. When a failure occurs in the heat dissipation component, the temperature of the motor will gradually increase, and this propagation path of temperature changes will be captured by the gated recurrent unit, thus generating a thermal imaging feature vector.

[0061] Step S123: Perform multi-scale sliding window segmentation on the energy consumption fluctuation parameter subsequence to generate short-term energy consumption trend segments and long-term energy consumption baseline segments, and input them into the parallel branches of the third time series feature encoder respectively. Extract short-term energy consumption mutation features through dilated convolutional kernels, and fuse long-term energy consumption baseline features through the self-attention mechanism to generate an energy consumption feature vector.

[0062] For the energy consumption fluctuation parameter subsequence, multi-scale sliding window segmentation is required. In the ventilation system, the short-term energy consumption trend segment can reflect the energy consumption change of the ventilator in a short period (such as during the start-up or shutdown process), and the long-term energy consumption baseline segment reflects the average energy consumption level of the ventilator over a longer period (such as a day or a week). Input these two segments into the parallel branches of the third time series feature encoder respectively. Short-term energy consumption mutation features can be extracted through dilated convolutional kernels. For example, when the ventilator suddenly encounters a large resistance, the short-term energy consumption will mutate. Fuse the long-term energy consumption baseline features through the self-attention mechanism to generate an energy consumption feature vector.

[0063] Step S124: Construct a cross-modal feature fusion module, splice the vibration feature vector, thermal imaging feature vector, and energy consumption feature vector in tensors, calculate the dynamic correlation weights between different modal features using the multi-head attention mechanism to generate a cross-modal correlation matrix, and output the fused device state feature tensor through feature channel weighted pooling.

[0064] In the ventilation system, the features of the ventilator in terms of vibration, temperature, and energy consumption are interrelated. For example, abnormal vibration may lead to an increase in energy consumption and may also be accompanied by a rise in temperature. Calculate the dynamic correlation weights between different modal features using the multi-head attention mechanism to generate a cross-modal correlation matrix. For example, it may be determined that the correlation weight of the vibration feature for the overall fault judgment is 0.3 at a certain moment, the weight of the thermal imaging feature is 0.4, and the weight of the energy consumption feature is 0.3. Then output the fused device state feature tensor through feature channel weighted pooling.

[0065] Step S125: Input the device state feature tensor into the spatio-temporal feature aggregation module, combine the device component topology structure diagram data, aggregate the state propagation features of adjacent components through the graph convolutional network, and adopt a timestamp alignment strategy to fuse the difference gradient between the historical state feature and the real-time state feature to generate a spatio-temporal aggregation feature matrix.

[0066] There are physical connection relationships among components such as the impeller, motor, and housing of the ventilator. The state propagation characteristics of adjacent components are aggregated through a graph convolutional network. For example, a fault in the impeller may be transmitted to the motor through the shaft, affecting the operating state of the motor, and such state propagation characteristics between components will be aggregated. At the same time, a timestamp alignment strategy is adopted to fuse the differential gradients of historical state features and real-time state features. For instance, the operating state of the ventilator in the past week is compared with the current real-time state, and considering the influence of time factors on the component state, a spatio-temporal aggregation feature matrix is generated.

[0067] Step S126, based on the spatio-temporal aggregation feature matrix, construct a dynamic weight assignment layer, calculate the contribution degree of each feature dimension in the preset fault mode discrimination space through a learnable parameter matrix, generate a feature weight vector, and adaptively filter low-contribution features using a soft threshold function, and output an optimized high-dimensional fault feature map.

[0068] In the fault mode discrimination space of the ventilation system, different feature dimensions have different contribution degrees. For example, the vibration feature of the impeller has a greater contribution to judging the fault of the impeller itself, while the temperature feature of the motor has a greater contribution to judging the fault of the motor. Calculate the contribution degree of each feature dimension in the preset fault mode discrimination space through a learnable parameter matrix, and generate a feature weight vector. Use a soft threshold function to adaptively filter low-contribution features. For example, if the contribution degree of a certain feature to fault judgment is lower than a certain threshold, it will be filtered out, so as to output an optimized high-dimensional fault feature map.

[0069] Step S127, input the high-dimensional fault feature map into the anomaly mode detection module, reconstruct the device state reference distribution through a deep autoencoder, calculate the abnormal area in the reconstruction error matrix that exceeds the preset threshold, generate a potential anomaly mode vector, and perform a time window sliding match on the anomaly mode in combination with the device operation context information, and output a set of candidate fault modes.

[0070] Assume that the state reference distribution when the ventilator is operating normally is pre-learned through a deep autoencoder. When the ventilator shows anomalies, such as wear of the impeller or a fault in the motor, calculate the abnormal area in the reconstruction error matrix that exceeds the preset threshold, and generate a potential anomaly mode vector. For example, if the wear of the impeller causes the vibration frequency and amplitude to exceed the normal range, corresponding abnormal areas will appear in the reconstruction error matrix. Perform a time window sliding match on the anomaly mode in combination with the ventilator operation context information. Considering the operating characteristics of the ventilator in different seasons and different operating durations, analyze the anomaly mode, and output a set of candidate fault modes, which may include candidate fault modes such as impeller wear, motor overheating, and ventilation duct blockage.

[0071] Step S128: Perform similarity measurement on each candidate fault mode in the candidate fault mode set and the standard fault feature map in the expert knowledge base to generate an initial matching probability, and adjust the initial matching probability according to the cumulative occurrence frequency of the same type of fault mode in the device operation historical data to generate a dynamically weighted fault mode probability distribution.

[0072] For example, for the candidate fault mode of impeller wear, compare it with the standard fault feature map of impeller wear in the expert knowledge base. The map stipulates features such as the vibration frequency range and temperature change range during impeller wear, and perform matching according to the actual situation to generate an initial matching probability. Then adjust the initial matching probability according to the cumulative occurrence frequency of the same type of fault mode (impeller wear) in the historical operation data of the ventilator. If the frequency of impeller wear faults is relatively high in the past operation history, then appropriately increase the matching probability of the current candidate fault mode to generate a dynamically weighted fault mode probability distribution.

[0073] Step S129: Construct a belief propagation network, and based on the physical connection relationship and fault propagation path between device components, perform belief diffusion calculation on the fault mode probability distribution. By iteratively updating the abnormal belief weights of associated components, generate a belief propagation matrix containing the fault influence factors between components, perform tensor fusion on the belief propagation matrix and the dynamically weighted fault mode probability distribution, and generate the multi-dimensional fault feature distribution matrix through the normalized exponential function, where the row dimension of the multi-dimensional fault feature distribution matrix represents the probability weight distribution on the preset diagnosis dimension, and the column dimension represents the abnormal belief gradient of associated components.

[0074] For example, impeller wear may cause a decrease in ventilation volume, which in turn affects the performance of the entire ventilation system, and perform belief diffusion calculation on the fault mode probability distribution. By iteratively updating the abnormal belief weights of associated components (such as motors, ventilation ducts, etc.), generate a belief propagation matrix containing the fault influence factors between components.

[0075] Finally, in the multi-dimensional fault feature distribution matrix, the row dimension represents the probability weight distribution on the preset diagnosis dimension, such as the probability weights of different types of faults (such as mechanical faults, electrical faults, etc.), and the column dimension represents the abnormal belief gradient of associated components (impeller, motor, ventilation duct, etc.). This matrix can comprehensively reflect the potential fault situation of the ventilation system and provide an important basis for subsequent fault diagnosis and maintenance.

[0076] In a possible implementation manner, step S140 includes:

[0077] Step S141: Calculate the reliability loss value of each verified diagnostic unit in the standard diagnostic report and the error correction value of each unverified diagnostic unit in the emergency diagnostic report. The reliability loss value is the cross-validation result of the equipment status parameters within the diagnostic unit on the fault mode discrimination boundary.

[0078] Step S142: Dynamically weight and fuse the mean reliability loss of the verified diagnostic units and the mean error correction of the unverified diagnostic units to generate an optimized target component for diagnostic accuracy.

[0079] Suppose there are 3 verified diagnostic units with reliability loss values of 0.1, 0.15, and 0.08 respectively, and their mean is (0.1 + 0.15 + 0.08) / 3 = 0.11. There are 2 unverified diagnostic units with error correction weights of 0.3 and 0.25 respectively, and their mean is (0.3 + 0.25) / 2 = 0.275. These two means are fused through a dynamic weighting coefficient (this coefficient may be preset according to different system conditions or adjusted during the training process). For example, if the weighting coefficient for the mean of the verified diagnostic units is set to 0.6 and the weighting coefficient for the mean of the unverified diagnostic units is set to 0.4, then the optimized target component for diagnostic accuracy is 0.11×0.6 + 0.275×0.4 = 0.172.

[0080] Step S143: Extract the high-order representation vectors of the standard diagnostic report and the emergency diagnostic report in the equipment feature latent space, calculate the distribution difference degree between the two in the equipment health status discrimination space, and generate an optimized target component for feature alignment.

[0081] For the standard diagnostic report of an air conditioning system, it is transformed into a benchmark vector of equipment health status through a deep feature extraction network. This vector contains the life cycle characteristics and compound fault correlation patterns of various components of the air conditioner (such as compressors, evaporators, condensers, etc.). For example, the remaining expected service life of the compressor, and the correlation between compressor faults and faults of other components such as evaporators and condensers. Map the emergency diagnostic report to a real-time operation and maintenance response feature vector, which contains the abnormal propagation path and the impact factor of emergency measures. For example, if refrigerant leakage is suspected, then the abnormal propagation path may be that the reduction of refrigerant leads to a decrease in cooling effect, which may further affect the temperature distribution of the evaporator, and the temporary measures taken on site (such as checking whether there are leakage points in the refrigerant pipeline) are the impact factors of emergency measures.

[0082] Suppose the benchmark vector of equipment health status is vector A and the real-time operation and maintenance response feature vector is vector B. Calculate the cosine similarity between the two according to the dot product formula of vectors. For example, if A = (a1, a2, a3) and B = (b1, b2, b3), then the cosine similarity = (a1×b1 + a2×b2 + a3×b3) / (sqrt(a12 + a22 + a32 )×sqrt(b12 + b22 + b3 2 ))。Meanwhile, the cosine similarity is dynamically scaled based on the device operating environment parameters (such as the temperature and humidity in the room where the air conditioner is located, the usage frequency of the air conditioner, etc.). For example, if the room temperature is too high, it may affect the calculation of this cosine similarity, and it is adjusted according to the preset rules.

[0083] Optimize the feature encoder of the fault prediction neural network through the adversarial training mechanism to make the cosine similarity reach the preset stability threshold. During the training process, the parameters of the feature encoder are continuously adjusted to increase the value of the cosine similarity until the preset stability threshold is reached. This process makes the distribution of the high-order representation vectors of the standard diagnostic report and the emergency diagnostic report in the device feature latent space closer in the fault traceability space, thereby generating the feature alignment optimization target component.

[0084] Step S144, non-linearly couple the diagnostic accuracy optimization target component and the feature alignment optimization target component to generate the composite training objective function.

[0085] There are various ways for this non-linear coupling. For example, using a complex functional relationship, such as setting the diagnostic accuracy optimization target component as x and the feature alignment optimization target component as y, the composite training objective function can be in the form of f(x, y) = x 2 + y2 + 0.5xy, etc. This composite training objective function comprehensively considers the factors of diagnostic accuracy and feature alignment, and can comprehensively guide the training process of the fault prediction neural network, making it more accurate and reliable when dealing with the fault prediction of the air conditioning system.

[0086] In a possible implementation manner, step S141 includes:

[0087] Step S1411, for the device status discrimination node in each diagnostic unit, obtain the credibility distribution of the corresponding fault mode of the device status discrimination node in the multi-dimensional fault feature distribution matrix.

[0088] In this embodiment, taking the air conditioning system in a building as an example, for the standard diagnostic report of the air conditioning system, the verification and diagnosis unit therein contains accurate judgment bases for various fault situations. In each verification and diagnosis unit, there are multiple equipment state discrimination nodes. For example, for the verification and diagnosis unit of the air conditioning compressor fault, the equipment state discrimination nodes may include parameters such as the suction pressure, discharge pressure, and operating current of the compressor. For each equipment state discrimination node, obtain the confidence distribution corresponding to the fault mode (such as compressor fault) in the multi-dimensional fault feature distribution matrix. Taking the suction pressure discrimination node of the compressor as an example, the multi-dimensional fault feature distribution matrix will give the confidence when the suction pressure is in different numerical ranges under the compressor fault mode.

[0089] Step S1412, based on the dynamic deviation degree between the equipment operation baseline parameters and the real-time monitoring parameters, construct the confidence attenuation function of each discrimination node in the diagnosis unit.

[0090] The air conditioning system has a set of equipment operation baseline parameters during normal operation. For example, the suction pressure of the compressor has a standard range under normal working conditions. If the real-time monitored suction pressure deviates from this standard range, construct the confidence attenuation function according to the deviation degree. Assume that the baseline range of the suction pressure is 0.4 - 0.6 MPa. When the real-time monitored value is 0.3 MPa, the deviation degree is 0.1 MPa. According to the preset functional relationship, the confidence of this discrimination node may be reduced from the initial 0.9 to 0.6.

[0091] Step S1413, calculate the pedigree matching loss between the confidence distribution of each discrimination node in the verification and diagnosis unit and the preset expert-annotated fault feature map.

[0092] The expert-annotated fault feature map details the features that each equipment state discrimination node should have when the compressor fails. For example, the fault feature map stipulates that when the compressor fails, the confidence that the suction pressure is lower than 0.35 MPa is 0.8. If the confidence calculated according to the multi-dimensional fault feature distribution matrix that the suction pressure is lower than 0.35 MPa is 0.6, then there will be a pedigree matching loss, which reflects the difference degree between the actual confidence distribution and the expert annotation.

[0093] Step S1414, count the misjudgment frequency of the abnormal speculation nodes in the historical operation and maintenance cases in the un-verified diagnosis unit, and generate the dynamic error correction weight.

[0094] For the unverified diagnosis units in the emergency diagnosis report, for example, when the air conditioner suddenly has poor cooling effect, the on-site operation and maintenance personnel initially judge that it may be refrigerant leakage or evaporator frosting, but these are only unverified abnormal speculations. Count the misjudgment frequencies of these abnormal speculation nodes (such as the refrigerant leakage speculation node, the evaporator frosting speculation node) in historical operation and maintenance cases. Suppose that in the past 100 similar initial speculations, the number of times the refrigerant leakage speculation was confirmed to be wrong was 20 times, then the misjudgment frequency of this abnormal speculation node is 0.2. Generate dynamic error correction weights based on these misjudgment frequencies. The higher the misjudgment frequency, the lower the weight.

[0095] In a possible implementation manner, the method further includes:

[0096] Real-time access the maintenance decision records of the building equipment management system, and extract the optimized diagnosis segments that have been verified and the temporary diagnosis segments to be reviewed.

[0097] During the training process, dynamically adjust the sample weights of the standard diagnosis report and the emergency diagnosis report. The weight coefficient of the optimized diagnosis segment increases with its effectiveness evaluation value in actual operation and maintenance.

[0098] In this embodiment, the process is described by taking the air conditioning system in a building as an example. In the building equipment management system, the maintenance decision records of the air conditioning system are continuously accessed in real time. These maintenance decision records contain a lot of relevant information about air conditioning fault diagnosis and repair.

[0099] From these records, the optimized diagnosis segments that have been verified can be accurately extracted. For example, in the previous case where the air conditioner had poor cooling effect, after a series of detailed inspections and analyses, it was determined that the reduction in cooling efficiency was caused by the blockage of the air conditioner filter. This complete process from fault detection to accurate positioning belongs to the optimized diagnosis segment that has been verified. At the same time, the temporary diagnosis segments to be reviewed can also be extracted from them. For example, when the air conditioner makes abnormal noises, it is initially judged that it may be due to loose fan blades, but the final confirmation has not been made yet. This is the temporary diagnosis segment to be reviewed.

[0100] During the training process of the fault prediction neural network, it is necessary to dynamically adjust the sample weights of the standard diagnostic report and the emergency diagnostic report. For the optimized diagnostic fragment, its weight coefficient is closely related to the effectiveness evaluation value in actual operation and maintenance and shows an increasing trend. Still taking the optimized diagnostic fragment of air conditioner filter clogging as an example, if in actual operation and maintenance, the diagnostic and repair strategies corresponding to this diagnostic fragment are successfully applied multiple times, then its weight coefficient in the standard diagnostic report will continue to increase. This means that in subsequent training processes, the information contained in this optimized diagnostic fragment will be given higher importance. For the content in the emergency diagnostic report, due to the uncertainty of the temporary diagnostic fragments to be reviewed, the adjustment of its weight coefficient is relatively complex. For example, for the temporary diagnostic fragment of loose fan blades to be reviewed, if it is verified to be an incorrect diagnosis later, its weight in the emergency diagnostic report may decrease; if it is verified to be correct, its weight may be gradually adjusted according to the actual operation and maintenance situation to ensure that accurate diagnostic information can be fully utilized during the training process, improving the accuracy and reliability of the fault prediction neural network for air conditioner system fault diagnosis.

[0101] In a possible implementation manner, step S143 may include:

[0102] Step S1431, converting the standard diagnostic report into a device health status reference vector through a deep feature extraction network, where the device health status reference vector includes component life cycle characteristics and compound fault association patterns.

[0103] In this embodiment, taking the elevator system in a building as an example, for the standard diagnostic report of the elevator system, it includes detailed fault diagnosis information and repair strategies for each component of the elevator. The deep feature extraction network will deeply mine this information to generate a device health status reference vector. The component life cycle characteristics in this vector cover the service life of key components of the elevator, such as the expected remaining service life of the traction machine and the wear life cycle of the car guide rails. The compound fault association pattern reflects the internal relationship between faults of different components. For example, when a fault occurs in the traction machine, it may affect the running stability of the car, and then may cause abnormal triggering of safety devices. This association relationship will be incorporated into the device health status reference vector.

[0104] Step S1432, mapping the emergency diagnostic report into a real-time operation and maintenance response feature vector, where the real-time operation and maintenance response feature vector includes an abnormal propagation path and an emergency measure impact factor.

[0105] For example, during the operation of an elevator, an emergency diagnosis report is generated when an unexpected situation occurs. For example, when the elevator suddenly stops running, the on-site operation and maintenance personnel obtain an emergency diagnosis report after a preliminary inspection. The information in this report is mapped into a real-time operation and maintenance response feature vector. The abnormal propagation path therein describes the possible propagation process of the fault. For example, if abnormal shaking of the car is found, it may be due to local damage to the car guide rail, and this damage may further affect the safety distance between the car and the hoistway wall, which is the abnormal propagation path. The emergency measure impact factor reflects the impact of the temporary measures taken by the on-site operation and maintenance personnel on the overall state of the elevator. For example, the manual brake release operation taken by the on-site operation and maintenance personnel to move the car to the nearest floor to evacuate passengers, and the influencing factors of this operation on the subsequent state of the elevator will be included in the real-time operation and maintenance response feature vector.

[0106] Step S1433: Calculate the cosine similarity between the device health status reference vector and the real-time operation and maintenance response feature vector in the fault tracing space, and dynamically scale the cosine similarity based on the device operation environment parameters.

[0107] For example, in an elevator system, both the device health status reference vector and the real-time operation and maintenance response feature vector are in a specific vector space, namely the fault tracing space. Suppose the device health status reference vector is vector A, which contains elements such as the traction machine life cycle eigenvalue a1 and the car guide rail wear life cycle eigenvalue a2; the real-time operation and maintenance response feature vector is vector B, which contains elements such as the abnormal propagation path eigenvalue b1 of car shaking and the emergency measure impact factor eigenvalue b2 of manual brake release. According to the cosine similarity formula, calculate the cosine similarity between these two vectors. The specific calculation process is as follows: First, calculate the dot product of vector A and vector B, that is, a1×b1 + a2×b2 + …, then calculate the magnitudes of vector A and vector B respectively. For example, the magnitude of vector A is sqrt(a12 + a22 + …), and the magnitude of vector B is sqrt(b12 + b22 + …). Finally, divide the dot product by the product of the two magnitudes to obtain the cosine similarity. At the same time, dynamically scale the cosine similarity based on the device operation environment parameters. The operation environment parameters of the elevator include the usage frequency of the elevator, the height of the building where it is located, the running floor range, etc. For example, if the usage frequency of the elevator is high, then when calculating the cosine similarity, the weights of some wear-related features may increase, thereby adjusting the cosine similarity. If the building height is high, it may affect the weights of safety device-related features in the calculation of the cosine similarity, so as to achieve dynamic scaling based on the device operation environment parameters.

[0108] Step S1434: Optimize the feature encoder of the fault prediction neural network through an adversarial training mechanism to make the cosine similarity reach a preset stability threshold.

[0109] Finally, in the fault prediction neural network of the elevator system, the feature encoder is responsible for extracting features from various input data. During this process, an adversarial training mechanism is adopted to optimize the feature encoder. The adversarial training mechanism involves two networks, one is the generator network (here the feature encoder can be regarded as part of the generator network), and the other is the discriminator network. The goal of the generator network is to generate feature representations that can make the cosine similarity close to a preset stability threshold, while the discriminator network attempts to distinguish whether the generated feature representations meet the requirements. For example, the initially calculated cosine similarity is 0.6, and the preset stability threshold is 0.8. Under the action of the generator network, the feature encoder adjusts the feature extraction methods for the standard diagnostic report and the emergency diagnostic report, causing changes in the feature representations of the device health state reference vector and the real-time operation and maintenance response feature vector. The discriminator network judges this change. If it believes that the requirements have not been met, it feeds back to the generator network for further adjustment. With continuous iterative training, the feature encoder is gradually optimized, and the cosine similarity continuously approaches the preset stability threshold, thereby achieving the purpose of optimizing the fault prediction neural network, generating a feature alignment optimization target component, which can reflect the alignment degree of the standard diagnostic report and the emergency diagnostic report in the device feature latent space, and helps to improve the accuracy and reliability of the entire fault prediction system for elevator system fault diagnosis.

[0110] In a possible implementation manner, the method further includes:

[0111] Step A110, constructing a device operation state transfer learning framework, and using the device feature template under historical normal working conditions as a reference anchor point in the device operation state transfer learning framework.

[0112] In this embodiment, taking the ventilation system in a building as an example, for the ventilation system in a building, the construction of the device operation state transfer learning framework is to better utilize historical data to optimize fault prediction. The ventilation system includes components such as fans, air ducts, and filters. Under historical normal working conditions, by collecting long-term operation data of the ventilation system, a device feature template is constructed. This device feature template contains information in multiple aspects. For example, the rotation speed range of the fan during normal operation, the air pressure stability range in the air duct, and the resistance range of the filter under normal filtration efficiency. These data form a comprehensive device feature template and are used as a reference anchor point. This is like determining an origin in a coordinate system, and subsequent calculations and analyses will be based on this origin. For example, the rotation speed of the fan during normal operation is stable at 1000 - 1200 revolutions per minute, and this rotation speed range is part of the device feature template and serves as an important reference basis in the transfer learning framework.

[0113] Step A120: Calculate the deviation trajectory of the current predicted feature from the reference anchor point in the degradation trend space in each training iteration.

[0114] During the training process of the fault prediction neural network for the ventilation system, the current predicted feature is generated in each iteration. These predicted features reflect the current estimated operating state of the ventilation system. For example, in a certain training iteration, the predicted fan speed is 1100 revolutions per minute, the air duct air pressure is 0.8 standard atmospheric pressure, and the filter resistance is 10 pascals. The fan speed range in the reference anchor point is 1000 - 1200 revolutions per minute, the standard value of the air duct air pressure is 1 standard atmospheric pressure, and the normal range of the filter resistance is 8 - 12 pascals. In the degradation trend space, calculate the deviation trajectory of these predicted features from the reference anchor point. For the fan speed, the deviation degree may be small, but the deviation of the air duct air pressure indicates that there may be a blockage in the air duct or a decrease in the fan efficiency. This deviation is not a simple numerical difference, but a deviation in the degradation trend space considering the overall operating logic of the ventilation system and the interrelationships of components. For example, if the fan speed gradually decreases and the air duct air pressure continuously deviates from the standard value in multiple consecutive iterations, then this deviation trajectory reflects the trend that the ventilation system may be developing towards a fault. By calculating the deviation trajectories of the predicted features of multiple components from the reference anchor point, the change in the operating state of the ventilation system can be comprehensively understood.

[0115] Step A130: Generate a regularization constraint term for the model parameters of the fault prediction neural network based on the dynamic change rate of the deviation trajectory, and incorporate the regularization constraint term for the model parameters into the composite training objective function.

[0116] Generate a model parameter regularization constraint term based on the dynamic change rate of the deviation trajectory of the ventilation system. For example, the deviation trajectory of the fan speed shows a change rate of decreasing by 50 revolutions per minute every 10 iterations within a certain period of time, and the deviation trajectory of the air duct air pressure shows a change rate of decreasing by 0.05 standard atmospheric pressure every 10 iterations. Based on these dynamic change rates, the constraints on the model parameters of the fault prediction neural network can be determined. If the change rate is too fast, it indicates that the model may be too sensitive or have deviations to the state changes of the ventilation system, and the model parameters need to be constrained. This constraint term will be incorporated into the composite training objective function. The composite training objective function originally optimizes the fault prediction neural network by comprehensively considering various factors (such as relevant factors in the standard diagnostic report, emergency diagnostic report, etc.). After adding this model parameter regularization constraint term based on the dynamic change rate of the deviation trajectory, the model can pay more attention to the actual operation state change trend of the ventilation system during the training process, avoiding overfitting or incorrect prediction. For example, without this constraint term, the model may overly focus on the prediction error at a certain moment and ignore the overall degradation trend of the ventilation system. After adding the constraint term, the model can better adapt to the long-term operation state change of the ventilation system while ensuring the overall prediction accuracy.

[0117] In a possible implementation manner, the method further includes:

[0118] Step B110, configure an online knowledge distillation mechanism, and use the fault decision tree model constructed by domain experts as the teacher network.

[0119] In the fault prediction of the ventilation system, configuring an online knowledge distillation mechanism is to improve the performance and interpretability of the fault prediction neural network. Domain experts construct a fault decision tree model as the teacher network based on years of ventilation system operation and maintenance experience and theoretical knowledge. This fault decision tree model covers various possible fault situations of the ventilation system and the corresponding diagnostic paths. For example, at the top node of the fault decision tree, it may be whether the ventilation volume is normal. If the ventilation volume is abnormal, the next layer of nodes may be whether the fan is operating normally. If the fan is operating normally, the next layer of nodes may be whether the air duct is blocked, etc. This hierarchical fault decision tree model can accurately judge the fault type and location based on different symptoms of the ventilation system. Using it as the teacher network can provide valuable prior knowledge for the fault prediction neural network.

[0120] Step B120, extract the diagnostic path features of the teacher network in typical fault scenarios, and generate a device state discrimination rule strengthening signal.

[0121] When a typical fault scenario occurs in the ventilation system, such as a complete blockage of the air duct, the teacher network (fault decision tree model) will have a specific diagnostic path. Starting from the initial situation where the fan is running normally but the ventilation volume is zero, gradually judge along the nodes of the fault decision tree to determine the diagnostic path of the air duct blockage. This diagnostic path contains multiple features, such as the normal fan speed but the excessive air pressure difference at both ends of the air duct. Extract these diagnostic path features and convert them into device state discrimination rule reinforcement signals. This signal contains the rule information on how to accurately judge faults based on the states of various components of the ventilation system in this typical fault scenario of air duct blockage. For example, this signal may contain specific discrimination rules for determining air duct blockage based on parameters such as the power of the fan, the air pressure difference of the air duct, and the ventilation volume. These rules can enhance the discrimination ability of the fault prediction neural network in similar scenarios.

[0122] Step B130, inject the device state discrimination rule reinforcement signal into the decision layer of the fault prediction neural network through the attention fusion module to generate an interpretable diagnostic logic flow.

[0123] In the fault prediction neural network of the ventilation system, the attention fusion module plays a key role. It injects the device state discrimination rule reinforcement signal into the decision layer of the fault prediction neural network. For example, when the fault prediction neural network judges whether there is an air duct blockage fault in the ventilation system, the decision layer originally makes a judgment based on the internal parameters and algorithms of the neural network. Now, with the device state discrimination rule reinforcement signal injected by the attention fusion module, the decision layer will also consider the discrimination rules extracted from the teacher network (fault decision tree model). For example, when the decision layer judges the air duct blockage, it will not only consider the relationship between the air pressure difference and ventilation volume of the air duct learned by the neural network, but also refer to the reasonable ranges of the air pressure difference and ventilation volume of the air duct under different fan powers obtained from the teacher network and other discrimination rules. In this way, the generated diagnostic logic flow becomes interpretable. In practical applications, if an air duct blockage is diagnosed, not only can this conclusion be given, but also based on the injected discrimination rules, it can be explained which specific numerical values and relationships of parameters such as fan power, air duct air pressure difference, and ventilation volume the conclusion is based on, improving the credibility and understandability of fault prediction.

[0124] Among them, the specific implementation steps of the online knowledge distillation mechanism include:

[0125] Establish a dual-channel contrastive learning architecture in the device feature encoding stage, and the teacher network channel outputs the diagnostic decision path based on the physical model.

[0126] Calculate the feature similarity between the prediction result of the student network and the decision path of the teacher network at the fault mode discrimination node.

[0127] Construct a knowledge transfer loss function based on the similarity, and dynamically adjust the ability of the student network to capture the latent fault features of the device based on the knowledge transfer loss function.

[0128] In this embodiment, in the device feature encoding stage of the ventilation system, a dual-channel contrastive learning architecture is established. Among them, the teacher network channel outputs a diagnostic decision path based on the physical model. The physical model of the ventilation system includes knowledge such as the fluid mechanics principle of the fan, the air flow model of the air duct, and the filtration principle of the filter. For example, according to the physical model of the fan, there is a specific relationship between the air volume and air pressure of the fan at different speeds. The teacher network channel uses this physical model knowledge. When receiving the operation data of the ventilation system (such as fan speed, air duct air pressure, ventilation volume, etc.), it outputs a diagnostic decision path based on the physical model. For instance, if the fan speed is normal but the ventilation volume is low, according to the physical models of the fan and the air duct, the teacher network channel may judge along the diagnostic decision path of air duct blockage or filter blockage. This decision path is constructed based on physical principles and empirical rules and can provide an accurate basis for fault prediction.

[0129] When the student network (fault prediction neural network) conducts fault prediction on the ventilation system, it will obtain prediction results, such as predicting whether the air duct is blocked, whether the fan is faulty, etc. Calculate the feature similarity between the prediction result of the student network and the decision path of the teacher network at the fault mode discrimination node. Taking the fault mode discrimination node of air duct blockage as an example, the teacher network decision path may judge the possibility of air duct blockage based on parameters such as air duct pressure difference and ventilation volume, and the student network will also analyze and predict these parameters according to the model it has learned. Calculate the similarity between the two in these parameter features. For example, if the teacher network judges the possibility of air duct blockage to be 80% based on the air duct pressure difference being greater than 0.5 standard atmospheric pressure and the ventilation volume being lower than 80% of the normal level, and the student network judges the possibility of air duct blockage to be 70% based on the same parameters, calculate the feature similarity between the two at this fault mode discrimination node through a specific algorithm (such as cosine similarity algorithm or Euclidean distance algorithm, etc.).

[0130] Construct a knowledge transfer loss function based on the feature similarity between the predicted results of the student network calculated and the decision-making path of the teacher network at the fault mode discrimination node. This loss function reflects the degree of difference between the student network and the teacher network. If the feature similarity is high, it indicates that the predicted results of the student network are relatively close to the decision-making path of the teacher network, and the value of the loss function is small; on the contrary, the value of the loss function is large. Dynamically adjust the ability of the student network to capture the latent fault features of the device based on this knowledge transfer loss function. For example, if the value of the knowledge transfer loss function is large at the fault mode discrimination node of air duct blockage, it means that the student network may not fully consider some latent fault features (such as small structural deformations inside the air duct that affect the ventilation volume although they do not completely block the air duct) when judging air duct blockage. By adjusting the parameters of the student network, such as increasing the weight of the features related to the internal structure of the air duct, improve the ability of the student network to capture these latent fault features, so that the student network can more accurately judge the fault conditions of the ventilation system in subsequent predictions, especially those latent fault conditions that are not easily detected.

[0131] In a possible implementation manner, the method further includes:

[0132] Step C110, establish a device state evolution prediction module, and construct a multi-dimensional health index curve based on the intermediate feature output of the fault prediction neural network.

[0133] In this embodiment, taking the elevator system in a building as an example, for the elevator system, the device state evolution prediction module aims to predict the evolution of the health state of the elevator in advance. When the fault prediction neural network performs fault prediction on the elevator system, it will generate intermediate feature outputs. These intermediate features contain the operation state information of each component of the elevator. For example, for the traction machine of the elevator, the intermediate features may include its vibration frequency and temperature change rate during operation; for the car, it may include the sway amplitude of the car and the stability of the running speed. Construct a multi-dimensional health index curve based on these intermediate feature outputs. Taking the vibration frequency of the traction machine as an example, assuming that the vibration frequency is between 10-20 Hz during normal operation, if the vibration frequency gradually increases beyond this range, it will be reflected in the corresponding dimension of the multi-dimensional health index curve. At the same time, the sway amplitude of the car is also reflected in another dimension of the curve. If the sway amplitude of the car gradually increases from the normal 0-1 mm to 2-3 mm, it will also be recorded on the multi-dimensional health index curve. In this way, the multi-dimensional health index curve comprehensively combines the state information of each key component of the elevator to form an index that comprehensively reflects the health state of the elevator.

[0134] Step C120, synchronously optimize the shared parameters of the device state evolution prediction module and the diagnosis module during the training process, so that the mutation points of the multi-dimensional health index curve are spatially and temporally correlated with the abnormal patterns of the multi-dimensional fault feature distribution matrix.

[0135] During the training process of the elevator system, there are shared parameters between the equipment state evolution prediction module and the diagnosis module. For example, for the parameters used to judge the traction machine failure, they are used in the diagnosis module to determine whether there is a failure and also in the equipment state evolution prediction module to predict the evolution of the health state of the traction machine. Synchronously optimizing these shared parameters is to ensure the spatio-temporal correlation between the mutation points of the multi-dimensional health index curve and the abnormal patterns of the multi-dimensional fault feature distribution matrix. When the bearings inside the traction machine start to wear, in the multi-dimensional fault feature distribution matrix, the features related to bearing wear (such as abnormal vibration frequency, temperature rise, etc.) will show abnormal patterns. At the same time, in the multi-dimensional health index curve, due to the effect of the shared parameters, mutation points will also appear at the corresponding time points. This mutation point may be a sudden increase in the vibration frequency dimension or a rapid increase in the temperature dimension. By synchronously optimizing the shared parameters, this spatio-temporal correlation can accurately reflect the time when the elevator fault occurs and the corresponding state changes. For example, if at a certain moment, mutation points simultaneously appear in the car sway amplitude and the traction machine vibration frequency in the multi-dimensional health index curve, and abnormal patterns of the car guide rail and the traction machine also appear in the multi-dimensional fault feature distribution matrix, it can be accurately judged that these two components may have simultaneous failures or related failures near this time point.

[0136] Step C130, capturing the stage features of the equipment degradation process through a temporal convolutional network to generate the timing trigger conditions for preventive maintenance decisions.

[0137] The elevator system will experience different degradation stages during long-term operation. The temporal convolutional network analyzes the operation data of the elevator system to capture these degradation stage features. For example, in the early degradation stage, some components of the elevator may only show slight wear, which may be reflected in the operation data as minor fluctuations in the vibration frequency of the traction machine and occasional slight instability in the running speed of the car. The temporal convolutional network can identify these early degradation features. As time goes by, entering the mid-term degradation stage, more obvious features may appear, such as an increase in the frequency of the traction machine temperature rise and an increase in the car sway amplitude. The temporal convolutional network captures the changing trends of these stage features through convolutional operations on the operation data at different time points. Based on these captured stage features, the timing trigger conditions for preventive maintenance decisions are generated. For example, if the temporal convolutional network detects that the number of times the fluctuation of the traction machine vibration frequency exceeds the normal range within a continuous month exceeds a certain threshold, and the car sway amplitude also shows a gradually increasing trend, preventive maintenance decisions can be triggered, such as arranging inspections of the traction machine and calibration of the car guide rail, etc., to avoid more serious failures caused by further degradation of the elevator.

[0138] In a possible implementation manner, the method further includes:

[0139] Step D110: Configure a dynamic adversarial training strategy to generate an adversarial sample set simulating abnormal states of the device.

[0140] In an elevator system, to improve the robustness of the fault prediction neural network, a dynamic adversarial training strategy is configured. First, an adversarial sample set simulating abnormal states of the device is generated. For different components of the elevator, adversarial samples are generated by performing specific transformations on the normal operation data. Taking the traction machine as an example, the vibration frequency, temperature, and other data during normal operation are known. Abnormal states are simulated by adding targeted noise or performing specific numerical transformations on this data. For example, on the vibration frequency data of the traction machine, some periodic interference signals are added according to a certain algorithm, making the vibration frequency seem like the situation when a certain fault (such as rotor imbalance) occurs. For the operating speed data of the car, the speed anomaly when the car guide rail is worn or the braking system fails may be simulated by changing the slope of the data or adding random fluctuations. The generated adversarial sample set contains various data simulating abnormal states of the elevator and can be used for subsequent dynamic adversarial training.

[0141] Step D120: Alternately input the real device data stream and the adversarial samples in the adversarial sample set in each training iteration, and calculate the diagnostic robustness index of the fault prediction neural network in a composite perturbation environment.

[0142] In each training iteration of the elevator system fault prediction neural network, the real device data stream and the adversarial samples in the adversarial sample set are alternately input. The real device data stream contains normal and abnormal data during the actual operation of the elevator. When the real device data stream is input, the fault prediction neural network performs fault prediction according to the normal prediction process. When the adversarial samples are input, since the adversarial samples simulate abnormal states, the fault prediction neural network needs to accurately identify these abnormalities. In this composite perturbation environment (with both real data and adversarial samples simulating abnormalities), the diagnostic robustness index is calculated. This index reflects the stability and accuracy of the fault prediction neural network in the face of various real and simulated situations. For example, the diagnostic robustness index can be measured by calculating the correct prediction ratio of the neural network for real abnormal data and abnormal data in the adversarial samples. If among 100 real abnormal data, the neural network correctly predicts 90, and among 100 abnormal data in the adversarial samples, it correctly predicts 80, then the diagnostic robustness index can be comprehensively calculated based on these two ratios.

[0143] Step D130: Based on the diagnostic robustness index, adaptively adjust the weight allocation ratio of each optimization component in the composite training objective function to generate a neural network iteration mechanism with environmental adaptability.

[0144] For example, according to the calculated diagnostic robustness index, adaptively adjust the weight allocation ratio of each optimization component in the composite training objective function. The composite training objective function includes multiple optimization components, such as components related to the standard diagnostic report, components related to the emergency diagnostic report, and so on. If the diagnostic robustness index is low, it indicates that the fault prediction neural network has insufficient ability to handle abnormal situations in the current training state. At this time, it may be necessary to increase the weight of the optimization component related to improving diagnostic accuracy. For example, if it is found that the prediction of abnormal data in the adversarial sample is inaccurate, it may be necessary to increase the weight of the optimization component related to the accurate fault location information in the standard diagnostic report, so that the neural network can pay more attention to learning accurate fault location knowledge. By this way of adaptively adjusting the weight allocation ratio, an iterative mechanism of the neural network with environmental adaptability is generated. This mechanism can dynamically adjust the training direction according to the performance of the fault prediction neural network at different training stages and in the face of different data types (real data and adversarial samples), so that the neural network continuously adapts to the actual operating environment of the elevator system and improves its fault prediction ability in various situations.

[0145] Figure 2 The hardware structure diagram of the AI intelligent diagnosis system 100 based on the intelligent system for implementing the above-mentioned AI intelligent diagnosis method based on the intelligent system is shown, as Figure 2 shown, the AI intelligent diagnosis system 100 based on the intelligent system may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0146] The machine-readable storage medium 120 can store data and / or instructions. In some embodiments, the machine-readable storage medium 120 can store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 can store the data and / or instructions used by the AI intelligent diagnosis system 100 based on the intelligent system to execute or use to complete the exemplary methods described in the present invention.

[0147] In the specific implementation process, one or more processors 110 execute the computer-executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the AI intelligent diagnosis method based on the intelligent system as described in the above method embodiments. The processor 110, the machine-readable storage medium 120, and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the transceiver actions of the communication unit 140.

[0148] For the specific implementation process of the processor 110, reference can be made to the various method embodiments executed by the AI intelligent diagnosis system 100 based on the intelligent system above. The implementation principles and technical effects are similar, and will not be elaborated here in this embodiment.

[0149] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above AI intelligent diagnosis method based on an intelligent system is implemented.

[0150] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing or description thereof.

Claims

1. An AI intelligent diagnosis method based on an intelligent system, characterized in that: The method comprises: The operation data stream of the building electromechanical equipment is collected through the edge computing node, and the state timing encoding is performed on the operation data stream to generate an equipment state timing encoding sequence, wherein the operation data stream includes the vibration spectrum, thermal imaging characteristics and energy consumption fluctuation parameters collected by multi-source heterogeneous sensors; Inputting the equipment state time series coding sequence into a fault prediction neural network, and outputting a multi-dimensional fault feature distribution matrix, wherein the multi-dimensional fault feature distribution matrix includes probability weights of equipment potential failure modes on preset diagnostic dimensions and abnormal confidence levels of associated components; Obtaining a standard diagnostic report generated by an expert knowledge base and an emergency diagnostic report generated by on-site operation and maintenance records, wherein the standard diagnostic report contains accurate fault location information and a complex repair strategy, and the emergency diagnostic report contains a temporary disposal plan and unverified abnormal speculation; Constructing a composite training objective function based on the multi-dimensional fault feature distribution matrix, the verified diagnostic unit of the standard diagnostic report, the unverified diagnostic unit of the emergency diagnostic report, and the equipment operation feature latent space mapping; The parameter space of the fault prediction neural network is optimized by the back propagation algorithm, so that after the composite training objective function reaches the convergence threshold, a corresponding target fault prediction neural network is generated, and fault prediction and diagnosis are performed on any input operation data stream of building electromechanical equipment based on the target fault prediction neural network.

2. The AI ​​intelligent diagnosis method based on the intelligent system according to claim 1 is characterized in that: The step of constructing a composite training objective function based on the multi-dimensional fault feature distribution matrix, the verified diagnostic unit of the standard diagnostic report, the unverified diagnostic unit of the emergency diagnostic report, and the equipment operation feature latent space mapping includes: Calculating the reliability loss value of each verified diagnostic unit in the standard diagnostic report and the error correction value of each unverified diagnostic unit in the emergency diagnostic report, wherein the reliability loss value is a cross-validation result of the device state parameter in the diagnostic unit on the fault mode discrimination boundary; Dynamically weighting and fusing the mean value of the reliability loss of the verified diagnostic unit and the mean value of the error correction of the unverified diagnostic unit to generate a diagnostic accuracy optimization target component; Extracting high-order representation vectors of the standard diagnostic report and the emergency diagnostic report in the equipment feature latent space, calculating the distribution difference between the two in the equipment health status discrimination space, and generating feature alignment optimization target components; The diagnostic accuracy optimization target component and the feature alignment optimization target component are nonlinearly coupled to generate the composite training objective function.

3. The AI ​​intelligent diagnosis method based on the intelligent system according to claim 2 is characterized in that: The step of calculating the reliability loss value of each verified diagnostic unit in the standard diagnostic report and the error correction value of each unverified diagnostic unit in the emergency diagnostic report includes: For each device state determination node in each diagnosis unit, obtaining the credibility distribution of the corresponding fault mode of the device state determination node in the multi-dimensional fault feature distribution matrix; Based on the dynamic deviation between the equipment operation baseline parameters and the real-time monitoring parameters, the confidence attenuation function of each discriminant node in the diagnosis unit is constructed; Calculating the spectral matching loss between the credibility distribution of each discriminant node in the verification diagnosis unit and the fault feature map annotated by a preset expert; The misjudgment frequency of abnormal inference nodes in the unverified diagnosis unit in historical operation and maintenance cases is counted to generate a dynamic error correction weight.

4. The AI ​​intelligent diagnosis method based on the intelligent system according to claim 1 is characterized in that: The method further comprises: Access the maintenance decision records of the building equipment management system in real time to extract the optimized diagnosis segments that have been verified and the temporary diagnosis segments to be reviewed; The sample weights of the standard diagnosis report and the emergency diagnosis report are dynamically adjusted during the training process, and the weight coefficient of the optimized diagnosis segment increases with its effectiveness evaluation value in actual operation and maintenance.

5. The AI ​​intelligent diagnosis method based on intelligent system according to claim 2 is characterized in that: The high-order characterization vectors of the standard diagnosis report and the emergency diagnosis report in the equipment feature latent space are extracted, and the distribution difference between the two in the equipment health status discrimination space is calculated. The steps for generating the feature alignment optimization target component include: Converting the standard diagnostic report into a device health status reference vector through a deep feature extraction network, wherein the device health status reference vector includes component life cycle characteristics and composite fault association patterns; Mapping the emergency diagnosis report into a real-time operation and maintenance response feature vector, wherein the real-time operation and maintenance response feature vector includes an abnormal propagation path and an emergency measure influencing factor; Calculating the cosine similarity between the equipment health status reference vector and the real-time operation and maintenance response feature vector in the fault tracing space, and dynamically scaling the cosine similarity based on equipment operating environment parameters; The feature encoder of the fault prediction neural network is optimized through an adversarial training mechanism so that the cosine similarity reaches a preset stability threshold.

6. The AI ​​intelligent diagnosis method based on intelligent system according to claim 1 is characterized in that: The method further comprises: Constructing a device operation state transfer learning framework, in which a device feature template under historical normal working conditions is used as a reference anchor point; In each training iteration, calculating the deviation trajectory of the current prediction feature from the reference anchor point in the degradation trend space; A model parameter regularization constraint term of the fault prediction neural network is generated based on the dynamic change rate of the deviation trajectory, and the model parameter regularization constraint term is integrated into the composite training objective function.

7. The AI ​​intelligent diagnosis method based on intelligent system according to claim 1 is characterized in that: The method further comprises: Configure an online knowledge distillation mechanism and use the fault decision tree model built by domain experts as the teacher network; Extracting diagnostic path features of the teacher network in typical fault scenarios and generating device status discrimination rule reinforcement signals; The device state discrimination rule reinforcement signal is injected into the decision layer of the fault prediction neural network through an attention fusion module to generate an explanatory diagnostic logic flow; The specific implementation steps of the online knowledge distillation mechanism include: A dual-channel contrastive learning architecture is established in the device feature encoding stage, and the teacher network channel outputs a diagnostic decision path based on the physical model; Calculate the feature similarity between the student network prediction result and the teacher network decision path at the fault mode discrimination node; A knowledge transfer loss function is constructed based on the similarity, and the ability of the student network to capture hidden fault features of the equipment is dynamically adjusted based on the knowledge transfer loss function.

8. The AI ​​intelligent diagnosis method based on intelligent system according to claim 1 is characterized in that: The method further comprises: Establishing an equipment state evolution prediction module, and constructing a multi-dimensional health index curve based on the intermediate feature output of the fault prediction neural network; During the training process, the shared parameters of the equipment state evolution prediction module and the diagnosis module are synchronously optimized so that the mutation point of the multi-dimensional health index curve forms a spatiotemporal association with the abnormal mode of the multi-dimensional fault feature distribution matrix; The stage characteristics of the equipment degradation process are captured through a temporal convolutional network, and the timing trigger conditions for preventive maintenance decisions are generated.

9. The AI ​​intelligent diagnosis method based on intelligent system according to claim 1 is characterized in that: The method further comprises: Configure dynamic adversarial training strategies to generate adversarial sample sets that simulate abnormal device states; In each training iteration, real device data streams and adversarial samples in the adversarial sample set are alternately input, and a diagnostic robustness index of the fault prediction neural network in a composite disturbance environment is calculated; The weight distribution ratio of each optimization component in the composite training objective function is adaptively adjusted based on the diagnostic robustness index to generate a neural network iteration mechanism with environmental adaptability.

10. An AI intelligent diagnosis system based on an intelligent system, characterized in that: The AI ​​intelligent diagnosis system based on the intelligent system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the AI ​​intelligent diagnosis method based on the intelligent system as described in any one of claims 1 to 9 above.

Images