AI intelligent diagnosis method and system based on intelligent system

Abstract

The present invention provides an AI intelligent diagnosis method and system based on an intelligent system, which uses edge computing nodes to collect building electromechanical equipment operation data streams containing multi-source heterogeneous data such as vibration spectra, thermal imaging features, and energy consumption fluctuation parameters, and performs state time series coding to generate a coding sequence. This sequence is input into a fault prediction neural network to obtain a multi-dimensional fault feature distribution matrix. Then, the standard diagnostic report of the expert knowledge base and the emergency diagnostic report of the on-site operation and maintenance record are obtained. Then, based on the above results and the latent space mapping of the equipment operation characteristics, a composite training objective function is constructed, and the fault prediction neural network parameter space is optimized by the back propagation algorithm to generate a target fault prediction neural network. Finally, the network is used to perform fault prediction and diagnosis on the building electromechanical equipment operation data stream to achieve comprehensive, accurate, and intelligent fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular 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 to ensuring the normal use of buildings. Traditional fault diagnosis methods for building electromechanical equipment have many limitations and cannot meet the increasingly complex electromechanical system requirements of modern buildings.

[0003] In the early days, equipment failures were primarily detected through regular manual inspections. Operations and maintenance personnel relied on experience and simple tools to inspect equipment. However, this method was inefficient, making it difficult to monitor equipment operating status in real time. It also made it difficult to detect potential early-stage failures. Failures were often discovered only after they had already developed to a significant degree and were significantly impacting equipment operation. This could lead to increased equipment damage, increased repair costs, and impacts on building functions.

[0004] With technological advancements, sensor-based monitoring methods have emerged. These methods typically collect data from a single sensor type, focusing on a single dimension like vibration signals or temperature changes. These methods fail to fully reflect the complex operational status of equipment. Because failure modes in building mechanical and electrical equipment are diverse and interconnected, monitoring using a single data source makes it difficult to accurately diagnose the cause and location of a fault, making misdiagnosis or missed diagnosis more likely.

[0005] Later, some intelligent diagnostic technologies incorporated machine learning algorithms. However, most of these algorithms simply performed simple classification or regression analysis after collecting large amounts of data, failing to fully consider the temporal characteristics of equipment operating data and the inherent connections between different types of data. For example, simply concatenating various collected data into a model without effectively processing the data to uncover deeper insights, resulting in limited ability to identify complex fault patterns and low diagnostic accuracy.

[0006] At the same time, existing fault diagnosis methods lack knowledge utilization. They either rely solely on historical data to train models, lacking the expertise of domain experts, making them unable to provide effective diagnostic results for complex and rare faults. Alternatively, they rely solely on expert experience to develop diagnostic rules, making them inflexible and unadaptable to the ever-changing operating conditions and emerging fault types of equipment. Furthermore, in actual maintenance, the valuable real-time information and temporary handling experience contained in on-site maintenance records are not fully utilized, making it impossible to combine this information with traditional diagnostic knowledge to optimize the diagnostic process. Summary of the Invention

[0007] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides an AI intelligent diagnosis method based on an intelligent system, the method comprising:

[0008] The operating data stream of the building's electromechanical equipment is collected through edge computing nodes, and the operating data stream is state-sequentially encoded to generate an equipment state sequence. The operating data stream includes vibration spectra, thermal imaging features, and energy consumption fluctuation parameters collected by multi-source heterogeneous sensors.

[0009] 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 potential equipment failure modes in preset diagnostic dimensions and abnormal confidence levels of associated components;

[0010] Obtain standard diagnostic reports generated by the expert knowledge base and emergency diagnostic reports generated by on-site operation and maintenance records. The standard diagnostic reports contain precise fault location information and complex repair strategies, while the emergency diagnostic reports contain temporary disposal plans and unverified anomaly speculations.

[0011] Constructing a composite training objective function based on the multidimensional 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;

[0012] 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 based on the target fault prediction neural network, fault prediction diagnosis is performed on the operating data stream of any input building mechanical and electrical equipment.

[0013] On the other hand, an embodiment of the present invention also provides an AI intelligent diagnostic system based on an intelligent system, including a processor and a machine-readable storage medium, wherein 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, the embodiment of the present application collects operating data streams such as vibration spectra, thermal imaging characteristics and energy consumption fluctuation parameters of building mechanical and electrical equipment obtained by multi-source heterogeneous sensors, and generates equipment status timing coding sequences through status timing coding. Compared with the traditional diagnostic method of a single data source or simple splicing of data, it can more comprehensively and accurately capture the complex characteristics of the equipment's operating status, 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 potential failure modes of equipment.

[0015] By inputting the equipment status time series coding sequence into the fault prediction neural network, the output is a multi-dimensional fault feature distribution matrix containing the probability weights of the equipment's potential failure modes in the preset diagnostic dimensions and the confidence levels of the associated component abnormalities. This breaks through the limitations of traditional diagnostic methods that only provide simple fault classifications or rough fault locations, and presents fault features in a more refined and quantitative manner, providing diagnostic personnel with rich and valuable decision-making basis, helping to more accurately determine the fault type and impact range, and improving the reliability and accuracy of fault diagnosis.

[0016] By acquiring standard diagnostic reports generated by an expert knowledge base and emergency diagnostic reports generated by on-site operation and maintenance records, and mapping these reports with a multidimensional fault feature distribution matrix and the latent space of equipment operation features, a composite training objective function is constructed. This combines the deep knowledge of domain experts with actual on-site operation and maintenance experience, overcoming the shortcomings of relying solely on historical data or a single diagnostic knowledge source. The precise fault location information and composite repair strategies in the standard diagnostic reports provide the model with accurate diagnostic references. The temporary solutions and unverified anomaly speculations in the emergency diagnostic reports introduce the flexibility and real-time factors of real-world scenarios, enabling the model to be optimized and trained based on a comprehensive set of information, thereby generating a target fault prediction neural network that better meets actual needs, is more practical, and more adaptable.

[0017] A backpropagation algorithm is used to optimize the parameter space of the fault prediction neural network, generating a target fault prediction neural network after the composite training objective function reaches a convergence threshold. This allows the fault prediction neural network to automatically adjust its parameters during training to balance the influence of multiple information sources. This not only improves the model's ability to fit known data but also enhances its generalization capabilities for arbitrary input data streams of building mechanical and electrical equipment operation. Compared to traditional diagnostic methods with fixed model structures or single optimization objectives, the target fault prediction neural network generated by this method is more adaptable to fault prediction and diagnosis of different operating conditions and types of building mechanical and electrical equipment, significantly improving the versatility and applicability of the diagnostic method. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of the execution flow of the AI ​​intelligent diagnosis method based on the intelligent system provided by an embodiment of the present invention.

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

[0020] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1The figure is a flow chart 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 an intelligent system is introduced in detail below.

[0021] In step S110, the operating data stream of the building's electromechanical equipment is collected through the edge computing node, and the operating data stream is state-time encoded to generate an equipment state time-series encoding sequence. The operating data stream includes the vibration spectrum, thermal imaging characteristics, and energy consumption fluctuation parameters collected by multi-source heterogeneous sensors.

[0022] In this embodiment, a large commercial building houses numerous electromechanical devices, such as air conditioning systems, elevator systems, and ventilation systems. Edge computing nodes are deployed near these devices. For the air conditioning system, multiple heterogeneous sensors are put into operation. A vibration sensor collects the vibration spectrum of the air conditioning compressor, recording vibrations at a constant frequency during operation, for example, collecting vibration data 100 times per second. This data reflects the motion of the compressor's internal mechanical components. A thermal imaging sensor captures thermal imaging features of the air conditioning condenser and evaporator, capturing the temperature distribution on the device surface. For example, certain areas of the evaporator surface may have lower temperatures, while heat dissipation issues may result in localized high-temperature areas. An energy consumption sensor collects energy consumption fluctuation parameters for the air conditioning system. Since energy consumption fluctuates significantly in different operating modes (such as cooling, heating, and dehumidification), the sensor records energy consumption data at regular intervals (e.g., every 5 minutes).

[0023] The edge computing node receives these data from different sensors and forms an operating data stream. This operating data stream is then encoded in state time series. Taking the vibration spectrum data of an air-conditioning compressor as an example, the continuously collected vibration frequency values ​​are encoded in chronological order. Assuming that the vibration frequency is 50 Hz at one moment and 51 Hz at the next, these two time points and the corresponding frequency values ​​are converted into part of the device state time series coding sequence according to specific coding rules. Thermal imaging features and energy consumption fluctuation parameters are also encoded in a similar manner, ultimately generating a complete device state time series coding sequence that fully records the operating status information of the building's electromechanical equipment over a period of time.

[0024] Step S120: input the equipment status time series coding sequence into a fault prediction neural network, and output a multi-dimensional fault feature distribution matrix, which includes the probability weights of the equipment's potential failure modes on preset diagnostic dimensions and the abnormal confidence levels of associated components.

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

[0026] First, the device state timing code sequence is modally separated using a multi-channel timing decomposition module. For elevator systems, the multi-source heterogeneous sensor data it contains is also decomposed. For example, the vibration spectrum subsequence collected from the vibration sensor at the bottom of the car contains the vibration state code vector of the car during operation, with precisely aligned timestamps. At the same time, the thermal imaging feature subsequences of key components such as motors collected by thermal imaging sensors are also separated, where each code vector reflects the temperature distribution correlation characteristics of the component surface at the corresponding moment. There is also an energy consumption fluctuation parameter subsequence, which reflects the energy consumption changes of the elevator during different operating stages (such as ascent, descent, and parking).

[0027] The vibration spectrum subsequence is input into the first time series feature encoder. The elevator car vibration data is processed through a time series convolutional network to extract local vibration waveform features. For example, when the car guide rails are slightly worn, a specific frequency change will be reflected in the local vibration waveform. Furthermore, a bidirectional long short-term memory network is used to capture long-term dependencies in the vibration pattern. For example, the vibration pattern of an elevator may vary periodically depending on the time of day. These are converted into vibration feature vectors.

[0028] The thermal imaging feature subsequence is input into the second temporal feature encoder. For elevator motors, a spatial attention mechanism is used to extract spatially correlated features of the thermal distribution on the motor surface. For example, if a blockage in the motor's heat sink causes localized overheating, this spatially correlated feature is captured. A gated recurrent unit is then used to capture the dynamic propagation path of temperature changes. When a fault within the motor prevents heat dissipation, the temperature gradually rises. This dynamic temperature propagation process is converted into a thermal imaging feature vector.

[0029] A multi-scale sliding window segmentation is performed on the energy consumption fluctuation parameter subsequences. For elevator systems, the short-term energy consumption trend segment may reflect the energy consumption changes during a single stop and start, while the long-term energy consumption baseline segment reflects the average energy consumption level of the elevator over a longer period of time (such as a day). These segments are respectively input into the parallel branches of the third temporal feature encoder. The short-term energy consumption mutation features are extracted using the dilated convolution kernel. For example, when the elevator's braking system malfunctions, a sudden change in short-term energy consumption may occur. The long-term energy consumption baseline features are fused through the self-attention mechanism to generate an energy consumption feature vector.

[0030] A cross-modal feature fusion module is constructed to concatenate the vibration feature vector, thermal imaging feature vector, and energy consumption feature vector into tensors. For example, in an elevator system, the three interrelated features of car vibration, motor temperature, and overall energy consumption are interrelated. A multi-head attention mechanism is used to calculate the dynamic correlation weights between them, generating a cross-modal correlation matrix. Then, through weighted pooling of feature channels, the fused device state feature tensor is output.

[0031] The device state feature tensor is input into the spatiotemporal feature aggregation module and combined with the elevator component topology graph data. For example, if there are physical connections between components such as the car, traction motor, and guide rails, the state propagation features of adjacent components are aggregated using a graph convolutional network. A traction motor failure may affect the operating state of the car through connected components, and this state propagation feature is aggregated. A timestamp alignment strategy is also used to fuse the difference gradients between historical state features and real-time state features. For example, the elevator's operating state over the past week is compared with the current real-time state to generate a spatiotemporal aggregation feature matrix.

[0032] A dynamic weight allocation layer is constructed based on a spatiotemporal aggregation feature matrix. For elevator systems, different feature dimensions contribute differently to the preset fault mode discrimination space. For example, the vibration characteristics of the elevator car contribute significantly to the identification of elevator guide rail faults. The contribution of each feature dimension is calculated using a learnable parameter matrix to generate a feature weight vector. A soft threshold function is then used to adaptively filter low-contribution features, outputting an optimized high-dimensional fault feature map.

[0033] The high-dimensional fault feature map is input into the abnormal pattern detection module. Assuming that the baseline state distribution during normal elevator operation is pre-learned using a deep autoencoder, when an abnormality occurs, such as abnormal car sway, the abnormal regions exceeding the preset threshold in the reconstruction error matrix are calculated to generate a potential abnormal pattern vector. A time window sliding matching is then performed on the abnormal pattern, combining the elevator's operational context. For example, considering the elevator's operational characteristics when stopping at different floors, a set of candidate fault modes is output, which may include candidate fault modes such as car guide rail wear and traction motor failure.

[0034] Each candidate fault mode in the candidate fault mode set is similarly measured against the standard fault signature map in the expert knowledge base. For example, the candidate fault mode for car guide rail wear is compared with the features of guide rail wear in the standard fault signature map to generate an initial matching probability. The initial matching probability is adjusted based on the cumulative occurrence frequency of similar fault modes in the elevator's historical operation data. For example, if the frequency of guide rail wear faults is high historically, the matching probability of the current candidate fault mode is appropriately increased, generating a dynamically weighted fault mode probability distribution.

[0035] A confidence propagation network is constructed based on the physical connections between elevator components and the fault propagation paths. For example, wear on the elevator car guide rails may affect the smoothness of the elevator car, which in turn may affect the working status of the safety devices. A confidence diffusion calculation is performed on the probability distribution of the fault mode. By iteratively updating the abnormal confidence weights of the associated components, a confidence propagation matrix is ​​generated that includes the factors affecting the faults between the components.

[0036] The confidence propagation matrix is ​​tensor-fused with the dynamically weighted fault mode probability distribution, generating a multidimensional fault signature distribution matrix using a normalized exponential function. In this matrix, the row dimension represents the probability weight distribution along the pre-set diagnostic dimensions, such as the probability weights for different fault types (e.g., mechanical faults, electrical faults), while the column dimension represents the abnormality confidence gradient of associated components, such as the abnormality confidence of components like the elevator car and traction motor.

[0037] Step S130, 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 composite repair strategy, and the emergency diagnostic report contains a temporary disposal plan and unverified abnormal speculation.

[0038] Taking the elevator system as an example, the expert knowledge base is built upon years of elevator maintenance experience and theoretical research. When an elevator malfunction occurs, the expert knowledge base generates a standard diagnostic report based on previous fault pattern identification and analysis. For example, if the fault is determined to be car guide rail wear, precise fault location information may indicate wear on a specific section of the car guide rail (e.g., between the 5th and 10th floors). A comprehensive repair strategy may include replacing the worn guide rail components, adjusting the car's balance, and other steps.

[0039] On-site operation and maintenance records are emergency diagnostic reports recorded by on-site maintenance personnel in emergency situations. For example, if an elevator suddenly stops, on-site maintenance personnel may find the car is shaking slightly after a preliminary inspection. Based on their experience, they may make some temporary solutions, such as parking the elevator at the nearest floor and evacuating passengers. They then record unverified anomaly hypotheses in the emergency diagnostic report, such as a foreign object stuck in the car guide rails or a problem with the car's balance, but these hypotheses have not yet been further verified.

[0040] Step S140 , constructing a composite training objective function based on the multi-dimensional fault feature distribution matrix, the verified diagnostic units of the standard diagnostic report, the unverified diagnostic units of the emergency diagnostic report, and the equipment operation feature latent space mapping.

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

[0042] In the standard diagnostic report, each verification diagnostic unit includes the basis for determining a specific fault. For example, for the verification diagnostic unit for car guide rail wear, the device status judgment node includes information such as the degree of guide rail wear and the location of wear. The credibility distribution of the device status judgment node for the corresponding fault mode (car guide rail wear) is obtained from the multidimensional fault signature distribution matrix. Simultaneously, a confidence decay function is constructed for each judgment node within the diagnostic unit based on the dynamic deviation between the device's operating baseline parameters and the real-time monitoring parameters. For example, under normal circumstances, there is a baseline parameter for the wear rate of the car guide rails. If the actual monitored wear rate exceeds this baseline, the confidence decays based on this deviation. The spectral matching loss is calculated between the credibility distribution of each judgment node in the verification diagnostic unit and the fault signature map annotated by the preset expert. If the fault signature map specifies specific behaviors when the guide rail wear reaches a certain level, and the actual monitored situation deviates from this, spectral matching loss will occur.

[0043] For unverified diagnostic units in emergency diagnosis reports, such as those previously suspected of a foreign object stuck in the car guide rail or a car balance problem, the frequency of misjudgments for these abnormal speculation nodes in historical operation and maintenance cases is counted to generate dynamic error correction weights. If the misjudgment rate for similar speculations in the past is high, the weight will be lower.

[0044] The mean reliability loss of the verified diagnostic unit and the mean error correction of the unverified diagnostic unit are dynamically weighted and fused to generate the diagnostic accuracy optimization target component.

[0045] Then, high-order representation vectors of the standard and emergency diagnostic reports in the device feature latent space are extracted. Standard diagnostic reports are converted into a baseline device health state vector using a deep feature extraction network. This vector contains component lifecycle characteristics and complex fault correlation patterns. Examples include the expected lifespan of elevator car guide rails and the correlation between guide rail wear and other component failures. Emergency diagnostic reports are mapped into a real-time operation and maintenance response feature vector, which contains the anomaly propagation path and the influencing factors of emergency measures. For example, a problem with the elevator car guide rails may affect the smooth operation of the car, representing the anomaly propagation path, while the temporary measures taken on-site (such as stopping and evacuating passengers) represent the influencing factors of emergency measures. The cosine similarity between the baseline device health state vector and the real-time operation and maintenance response feature vector in the fault tracing space is calculated and dynamically scaled based on the device's operating environment parameters. For example, environmental parameters such as the building height and frequency of use of the elevator can affect the calculation of this cosine similarity. An adversarial training mechanism is used to optimize the feature encoder of the fault prediction neural network, ensuring that the cosine similarity reaches a preset stability threshold, thereby generating the target component for feature alignment optimization.

[0046] Finally, the diagnostic accuracy optimization objective component and the feature alignment optimization objective component are nonlinearly coupled to generate a composite training objective function.

[0047] Step S150, optimizing the parameter space of the fault prediction neural network through the back propagation algorithm so that the composite training objective function reaches the convergence threshold, generating a corresponding target fault prediction neural network, and performing fault prediction diagnosis on any input operation data stream of building electromechanical equipment based on the target fault prediction neural network.

[0048] Taking the elevator system as an example, after constructing a composite training objective function, the backpropagation algorithm is used to optimize the parameter space of the fault prediction neural network. During each iteration, the neural network's weights, biases, and other parameters are adjusted based on the value of the composite training objective function. For example, if the composite training objective function's value is large, indicating a significant deviation between the current neural network prediction and the actual situation (such as standard diagnostic reports and emergency diagnostic reports), the backpropagation algorithm is used to propagate the error from the output layer to the input layer, adjusting the parameters of the intermediate layers.

[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 point, the targeted fault prediction neural network can perform fault prediction and diagnosis for any input elevator system operational data stream, such as newly collected data including car vibration spectra, motor thermal imaging characteristics, and energy consumption fluctuation parameters. It processes the input operational data stream according to a previously trained model and outputs a multidimensional fault feature distribution matrix containing the probability weights of potential fault modes along pre-set diagnostic dimensions and the confidence levels of abnormalities for associated components. This allows the system to accurately determine whether the elevator is faulty, as well as the likely fault type and location.

[0051] Based on the above steps, the embodiment of the present application collects operating data streams such as vibration spectra, thermal imaging characteristics and energy consumption fluctuation parameters of building mechanical and electrical equipment obtained by multi-source heterogeneous sensors, and generates equipment status timing code sequences through status timing coding. Compared with the traditional diagnostic method of a single data source or simple splicing of data, it can more comprehensively and accurately capture the complex characteristics of the equipment's operating status, 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 potential failure modes of equipment.

[0052] By inputting the equipment status time series coding sequence into the fault prediction neural network, the output is a multi-dimensional fault feature distribution matrix containing the probability weights of the equipment's potential failure modes in the preset diagnostic dimensions and the confidence levels of the associated component abnormalities. This breaks through the limitations of traditional diagnostic methods that only provide simple fault classifications or rough fault locations, and presents fault features in a more refined and quantitative manner, providing diagnostic personnel with rich and valuable decision-making basis, helping to more accurately determine the fault type and impact range, and improving the reliability and accuracy of fault diagnosis.

[0053] By acquiring standard diagnostic reports generated by an expert knowledge base and emergency diagnostic reports generated by on-site operation and maintenance records, and mapping these reports with a multidimensional fault feature distribution matrix and the latent space of equipment operation features, a composite training objective function is constructed. This combines the deep knowledge of domain experts with actual on-site operation and maintenance experience, overcoming the shortcomings of relying solely on historical data or a single diagnostic knowledge source. The precise fault location information and composite repair strategies in the standard diagnostic reports provide the model with accurate diagnostic references. The temporary solutions and unverified anomaly speculations in the emergency diagnostic reports introduce the flexibility and real-time factors of real-world scenarios, enabling the model to be optimized and trained based on a comprehensive set of information, thereby generating a target fault prediction neural network that better meets actual needs, is more practical, and more adaptable.

[0054] A backpropagation algorithm is used to optimize the parameter space of the fault prediction neural network, generating a target fault prediction neural network after the composite training objective function reaches a convergence threshold. This allows the fault prediction neural network to automatically adjust its parameters during training to balance the influence of multiple information sources. This not only improves the model's ability to fit known data but also enhances its generalization capabilities for arbitrary input data streams of building mechanical and electrical equipment operation. Compared to traditional diagnostic methods with fixed model structures or single optimization objectives, the target fault prediction neural network generated by this method is more adaptable to fault prediction and diagnosis of different operating conditions and types of building mechanical and electrical equipment, significantly improving the versatility and applicability of the diagnostic method.

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

[0056] Step S121, modally separate the device state timing code sequence through a multi-channel timing decomposition module to generate a vibration spectrum subsequence, a thermal imaging feature subsequence, and an energy consumption fluctuation parameter subsequence corresponding to the multi-source heterogeneous sensor type, each subsequence containing a device state code vector aligned with the timestamp.

[0057] In this embodiment, taking the ventilation system in a building as an example, the ventilation system contains data collected by multiple sensors, and these data constitute a device state time series coding sequence. When performing modal separation, subsequences corresponding to the types of multi-source heterogeneous sensors are generated. For example, the vibration spectrum data collected by the vibration sensor installed on the ventilator is separated into a vibration spectrum subsequence. Each device state coding vector in this vibration spectrum subsequence carries an accurate timestamp, which fully 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 are separated into thermal imaging feature subsequences, in which the device state coding vector accurately reflects the thermal imaging features such as the temperature distribution at the corresponding time. In addition, the energy consumption fluctuation parameters of the ventilation system collected by the energy consumption sensor are separated into energy consumption fluctuation parameter subsequences, and their device state coding vectors are also arranged in chronological order, covering the energy consumption conditions at different time points.

[0058] In step S122, the vibration spectrum subsequence is input into a first temporal feature encoder. A temporal convolutional network is used to extract local vibration waveform features. A bidirectional long short-term memory network is used to capture the long-term dependencies of the vibration pattern, generating a vibration feature vector. Simultaneously, the thermal imaging feature subsequence is input into a second temporal feature encoder. A spatial attention mechanism is used to extract the spatial correlation features of the thermal distribution. A gated recurrent unit is used to capture the dynamic propagation path of temperature changes, generating a thermal imaging feature vector.

[0059] During fan operation, its vibration spectrum contains a wealth of information. A time-series convolutional network processes subsequences of the vibration spectrum to extract local vibration waveform features. For example, if a fan's blades exhibit a slight imbalance, the local vibration waveform will exhibit fluctuations of specific frequency and amplitude. Furthermore, a bidirectional long short-term memory network captures the long-term dependencies of the vibration pattern. The fan's operating load varies at different times of the day, and these load variations can lead to long-term variations in the vibration pattern. This network can capture these long-term dependencies and generate vibration feature vectors.

[0060] At the same time, the thermal imaging feature subsequence is input into the second temporal feature encoder. For the ventilator's motor and heat dissipation components, the spatial attention mechanism focuses on the thermal distribution in the thermal imaging feature subsequence and extracts the spatial correlation features of the thermal distribution. For example, if there is a local short circuit in the winding part of the motor, high temperature in a specific area may appear on the thermal image, and the spatial correlation features between this high-temperature area and the surrounding area will be extracted. Then, the gated recurrent unit captures the dynamic propagation path of the temperature change. When the heat dissipation component fails, the temperature of the motor will gradually increase. The propagation path of this temperature change will be captured by the gated recurrent unit, thereby generating a thermal imaging feature vector.

[0061] In step S123, a multi-scale sliding window segmentation is performed on the energy consumption fluctuation parameter subsequence to generate short-term energy consumption trend segments and long-term energy consumption baseline segments, which are respectively input into the parallel branches of the third temporal feature encoder. The short-term energy consumption mutation features are extracted by the dilated convolution kernel, and the long-term energy consumption baseline features are fused 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 changes of the ventilator in a short period of time (such as the start-up or shutdown process), while the long-term energy consumption baseline segment reflects the average energy consumption level of the ventilator over a longer period of time (such as a day or a week). These two segments are input into the parallel branches of the third temporal feature encoder respectively. The short-term energy consumption mutation feature can be extracted by expanding the convolution kernel. For example, when the ventilator suddenly encounters a large resistance, the short-term energy consumption will mutate. The long-term energy consumption baseline feature is fused through the self-attention mechanism to generate an energy consumption feature vector.

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

[0064] In ventilation systems, the vibration, temperature, and energy consumption characteristics of ventilators are interrelated. For example, abnormal vibration may lead to increased energy consumption, which may also be accompanied by an increase in temperature. A multi-head attention mechanism is used to calculate the dynamic correlation weights between different modal features and generate a cross-modal correlation matrix. For example, the correlation weight of the vibration feature for the overall fault diagnosis may be determined to be 0.3 at a certain moment, the weight of the thermal imaging feature may be 0.4, and the weight of the energy consumption feature may be 0.3. The fused device state feature tensor is then output through weighted pooling of the feature channels.

[0065] In step S125, the device state feature tensor is input into the spatiotemporal feature aggregation module, combined with the device component topology structure data, the state propagation features of adjacent components are aggregated through the graph convolutional network, and the timestamp alignment strategy is used to fuse the difference gradients of historical state features and real-time state features to generate a spatiotemporal aggregation feature matrix.

[0066] The impeller, motor, and casing of a ventilator are physically connected to each other. A graph convolutional network aggregates the state propagation features of adjacent components. For example, an impeller failure may be transmitted to the motor through the shaft, affecting its operating status. This inter-component state propagation feature is aggregated. Furthermore, a timestamp alignment strategy is used to fuse the difference gradients between historical state features and real-time state features. For example, the operating status of a ventilator over the past week is compared with its current real-time status. Taking into account the impact of time on component status, a spatiotemporal aggregation feature matrix is ​​generated.

[0067] In step S126, a dynamic weight allocation layer is constructed based on the spatiotemporal aggregation feature matrix, the contribution of each feature dimension in the preset fault mode discrimination space is calculated through a learnable parameter matrix, a feature weight vector is generated, and a soft threshold function is used to adaptively filter low-contribution features to output an optimized high-dimensional fault feature map.

[0068] In the ventilation system's fault mode discrimination space, different feature dimensions contribute differently. For example, the vibration characteristics of the impeller contribute more to the determination of impeller faults, while the temperature characteristics of the motor contribute more to the determination of motor faults. A learnable parameter matrix is ​​used to calculate the contribution of each feature dimension in the preset fault mode discrimination space and generate a feature weight vector. A soft threshold function is used to adaptively filter low-contribution features. For example, if a feature's contribution to fault diagnosis falls below a certain threshold, it is filtered out, resulting in an optimized high-dimensional fault feature map.

[0069] In step S127, the high-dimensional fault feature map is input into the abnormal pattern detection module, the equipment state baseline distribution is reconstructed through the deep autoencoder, the abnormal area exceeding the preset threshold in the reconstructed error matrix is ​​calculated, the potential abnormal pattern vector is generated, and the abnormal pattern is subjected to time window sliding matching in combination with the equipment operation context information to output a set of candidate fault patterns.

[0070] It is assumed that the state baseline distribution of the ventilator during normal operation is pre-learned using a deep autoencoder. When an abnormality occurs in the ventilator, such as impeller wear or motor failure, the abnormal regions exceeding the preset threshold in the reconstructed error matrix are calculated to generate a potential abnormal pattern vector. For example, if impeller wear causes the vibration frequency and amplitude to exceed the normal range, a corresponding abnormal region will appear in the reconstructed error matrix. The abnormal pattern is then subjected to time window sliding matching based on the ventilator's operating context. Taking into account the operating characteristics of the ventilator in different seasons and at different operating times, the abnormal pattern is analyzed and a set of candidate fault modes is output, which may include candidate fault modes such as impeller wear, motor overheating, and ventilation duct blockage.

[0071] In step S128, each candidate fault mode in the candidate fault mode set is measured for similarity with the standard fault feature map in the expert knowledge base to generate an initial matching probability. The initial matching probability is adjusted according to the cumulative occurrence frequency of similar fault modes in the equipment operation history data to generate a dynamically weighted fault mode probability distribution.

[0072] For example, candidate fault modes for impeller wear are compared with standard fault characteristic maps for impeller wear in the expert knowledge base. The maps specify characteristics such as the vibration frequency range and temperature variation range of impeller wear. Matching is performed based on actual conditions to generate an initial matching probability. The initial matching probability is then adjusted based on the cumulative occurrence frequency of similar fault modes (impeller wear) in the fan's historical operating data. If the frequency of impeller wear faults is high in past operating history, the matching probability of the current candidate fault mode is appropriately increased to generate a dynamically weighted fault mode probability distribution.

[0073] Step S129: construct a confidence propagation network, perform confidence diffusion calculation on the fault mode probability distribution based on the physical connection relationship and fault propagation path between equipment components, generate a confidence propagation matrix containing inter-component fault influencing factors by iteratively updating the abnormal confidence weights of associated components, perform tensor fusion on the confidence propagation matrix and the dynamically weighted fault mode probability distribution, and generate the multi-dimensional fault feature distribution matrix through a normalized exponential function, wherein the row dimension of the multi-dimensional fault feature distribution matrix represents the probability weight distribution on the preset diagnostic dimension, and the column dimension represents the abnormal confidence gradient of the associated components.

[0074] For example, impeller wear can lead to a decrease in ventilation volume, which in turn affects the performance of the entire ventilation system. This involves performing a confidence diffusion calculation on the probability distribution of failure modes. By iteratively updating the abnormal confidence weights of associated components (such as motors and ventilation ducts), a confidence propagation matrix is ​​generated that includes the factors affecting the failures between components.

[0075] Finally, in the multidimensional fault signature distribution matrix, the row dimension represents the probability weight distribution along the pre-set diagnostic dimensions, such as the probability weights for different types of faults (such as mechanical and electrical faults), while the column dimension represents the confidence gradient of abnormalities for associated components (such as the impeller, motor, and ventilation duct). This matrix comprehensively reflects the potential faults of the ventilation system, providing an important basis for subsequent fault diagnosis and maintenance.

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

[0077] Step S141, 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 status parameter in the diagnostic unit on the fault mode judgment boundary.

[0078] Step S142 , 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.

[0079] Assume there are three verified diagnostic units with reliability loss values ​​of 0.1, 0.15, and 0.08, respectively. Their mean is (0.1 + 0.15 + 0.08) / 3 = 0.11. There are two unverified diagnostic units with error correction weights of 0.3 and 0.25, respectively. Their mean is (0.3 + 0.25) / 2 = 0.275. These two means are fused using a dynamic weighting coefficient (this coefficient can be pre-set or adjusted during training depending on the system's circumstances). For example, setting the weighting coefficient for the verified diagnostic unit mean to 0.6 and the weighting coefficient for the unverified diagnostic unit mean to 0.4 results in a target component for diagnostic accuracy optimization of 0.11 × 0.6 + 0.275 × 0.4 = 0.172.

[0080] Step S143 , extracting high-order representation vectors of the standard diagnostic report and the emergency diagnostic report in the device feature latent space, calculating the distribution difference between the two in the device health status discrimination space, and generating a feature alignment optimization target component.

[0081] For the standard diagnostic report of the air-conditioning system, it is converted into a device health status benchmark vector through a deep feature extraction network. This vector contains the life cycle characteristics and complex fault association patterns of various air-conditioning components (such as compressors, evaporators, condensers, etc.). For example, the expected remaining service life of the compressor, and the correlation between compressor failures and failures of components such as evaporators and condensers. The emergency diagnostic report is mapped into a real-time operation and maintenance response feature vector. This vector contains the abnormal propagation path and the influencing factors of emergency measures. For example, if a refrigerant leak is suspected, the abnormal propagation path may be that the refrigerant is reduced, resulting in a decrease in the cooling effect, which may in turn affect the temperature distribution of the evaporator, and the temporary measures taken on site (such as checking whether there are leaks in the refrigerant pipeline) are the influencing factors of emergency measures.

[0082] Assume that the device health status baseline vector is vector A and the real-time operation and maintenance response feature vector is vector B. The cosine similarity between the two is calculated according to the dot product formula of the 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 )). At the same time, the cosine similarity is dynamically scaled based on the device's operating environment parameters (such as the temperature and humidity of the room where the air conditioner is located, the frequency of air conditioner use, etc.). For example, if the room temperature is too high, it may affect the calculation of the cosine similarity, and it will be adjusted according to pre-set rules.

[0083] The feature encoder of the fault prediction neural network is optimized through an adversarial training mechanism, ensuring that the cosine similarity reaches a preset stability threshold. During training, the feature encoder parameters are continuously adjusted to increase the cosine similarity until the threshold is reached. This process brings the high-order representation vectors of the standard and emergency diagnostic reports in the device feature latent space closer together in the fault tracing space, thereby generating the target component for feature alignment optimization.

[0084] Step S144 , performing nonlinear coupling on the diagnostic accuracy optimization target component and the feature alignment optimization target component to generate the composite training objective function.

[0085] The nonlinear coupling here can be implemented in many ways, such as using a complex functional relationship, such as setting the diagnostic accuracy optimization target component to x and the feature alignment optimization target component to y. The composite training objective function can be f(x, y) = x 2 +y2+0.5xy, etc. This composite training objective function comprehensively considers both diagnostic accuracy and feature alignment, and can comprehensively guide the training process of the fault prediction neural network, making it more accurate and reliable in processing air-conditioning system fault prediction.

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

[0087] Step S1411 : for each device state determination node in each diagnosis unit, obtain the credibility distribution of the fault mode corresponding to the device state determination 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 diagnosis unit therein contains the basis for accurate judgment of various fault conditions. In each verification diagnosis unit, there are multiple equipment status judgment nodes. For example, for the verification diagnosis unit of the air-conditioning compressor fault, the equipment status judgment node may include parameters such as the compressor's suction pressure, exhaust pressure, and operating current. For each equipment status judgment node, obtain the credibility distribution of its corresponding fault mode (such as compressor fault) in the multi-dimensional fault feature distribution matrix. Taking the compressor's suction pressure judgment node as an example, the multi-dimensional fault feature distribution matrix will give the credibility of the suction pressure in different numerical ranges under the compressor fault mode.

[0089] Step S1412: construct a confidence attenuation function for each discrimination node in the diagnosis unit based on the dynamic deviation between the device operation baseline parameter and the real-time monitoring parameter.

[0090] During normal operation, an air conditioning system maintains a set of baseline operating parameters. For example, the compressor's suction pressure falls within a standard range under normal operating conditions. If the real-time monitored suction pressure deviates from this standard range, a confidence decay function is constructed based on the degree of deviation. For example, if the baseline suction pressure range is 0.4-0.6 MPa and the real-time monitored value is 0.3 MPa, the deviation is 0.1 MPa. Based on the preset function, this may cause the confidence of the decision node to decrease from the initial 0.9 to 0.6.

[0091] Step S1413 , 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 the preset expert.

[0092] The expert-annotated fault signature map specifies the characteristics that each device state judgment node should possess when a compressor fails. For example, the fault signature map specifies that the confidence level for suction pressure below 0.35 MPa in the event of a compressor failure is 0.8. If the confidence level for suction pressure below 0.35 MPa calculated from the multidimensional fault signature distribution matrix is ​​0.6, a spectral matching loss will occur, reflecting the degree of discrepancy between the actual confidence distribution and the expert annotation.

[0093] Step S1414 , counting the misjudgment frequencies of abnormal inference nodes in the unverified diagnosis unit in historical operation and maintenance cases, and generating dynamic error correction weights.

[0094] For unverified diagnostic units in emergency diagnostic reports, such as when an air conditioner suddenly experiences poor cooling performance, on-site operations and maintenance personnel initially suspect a refrigerant leak or evaporator frost. However, these are unverified anomaly speculations. The frequency of misjudgments for these anomaly speculation nodes (such as the refrigerant leak and evaporator frost) in historical operations and maintenance cases is calculated. Assuming that the refrigerant leak speculation was proven incorrect 20 times out of the past 100 similar preliminary speculations, the misjudgment frequency for this anomaly speculation node is 0.2. Dynamic error correction weights are generated based on these misjudgment frequencies; higher misjudgment frequencies are associated with lower weights.

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

[0096] 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.

[0097] The sample weights of the standard diagnostic reports and the emergency diagnostic reports are dynamically adjusted during the training process, and the weight coefficients of the optimized diagnostic segments increase with their effectiveness evaluation values ​​in actual operation and maintenance.

[0098] In this example, the air conditioning system in a building is used as an example to illustrate this process. In the building equipment management system, maintenance decision records for the air conditioning system are continuously accessed in real time. These maintenance decision records contain a wealth of relevant information for air conditioning fault diagnosis and repair.

[0099] From these records, we can accurately extract verified optimized diagnostic segments. For example, in the case of poor air conditioning cooling effect, after a series of detailed inspections and analyses, it was determined that the cooling efficiency was reduced due to the clogging of the air conditioner filter. This complete process from fault detection to precise positioning belongs to the verified optimized diagnostic segment. At the same time, temporary diagnostic segments to be reviewed can also be extracted from it. For example, when the air conditioner makes abnormal noise, the preliminary judgment may be that the fan blades are loose, but final confirmation has not yet been made. This is the temporary diagnostic segment to be reviewed.

[0100] During the training process of the fault prediction neural network, the sample weights of standard diagnostic reports and emergency diagnostic reports need to be dynamically adjusted. For optimized diagnostic segments, their weight coefficients are closely correlated with their effectiveness evaluations in actual operations and show an increasing trend. Taking the optimized diagnostic segment for a clogged air conditioner filter as an example, if the diagnostic and repair strategies corresponding to this segment are successfully applied repeatedly in actual operations, the weight coefficient in the standard diagnostic report will continue to increase. This means that the information contained in this optimized diagnostic segment will be given greater importance during subsequent training. However, adjusting the weight coefficients of emergency diagnostic reports is more complex due to the uncertainty inherent in provisional diagnostic segments awaiting review. For example, if a provisional diagnostic segment for a loose fan blade is subsequently verified to be incorrect, its weight in the emergency diagnostic report may be reduced. If it is verified to be correct, its weight may be gradually adjusted based on actual operations to ensure that accurate diagnostic information is fully utilized during training, improving the accuracy and reliability of the fault prediction neural network for air conditioning system fault diagnosis.

[0101] In a possible implementation, 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. The device health status reference vector includes component life cycle characteristics and complex fault association patterns.

[0103] In this embodiment, taking the elevator system in a building as an example, the standard diagnostic report for the elevator system 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 benchmark vector. The component life cycle characteristics in this vector cover the service life of key elevator components, such as the expected remaining service life of the traction motor and the wear life cycle of the car guide rails. The composite fault association pattern reflects the inherent connection between the failures of different components. For example, when the traction motor fails, it may affect the operating stability of the car, which may lead to abnormal triggering of the safety device. This association relationship will be incorporated into the equipment health status benchmark vector.

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

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

[0106] Step S1433 , 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.

[0107] For example, in an elevator system, both the equipment health status baseline vector and the real-time operation and maintenance response feature vector reside in a specific vector space called the fault tracing space. Assume that the equipment health status baseline vector is vector A, whose elements include the traction motor life cycle eigenvalue a1 and the car guide rail wear life cycle eigenvalue a2; and the real-time operation and maintenance response feature vector is vector B, whose elements include the car sway abnormality propagation path eigenvalue b1 and the manual brake release emergency response factor eigenvalue b2. The cosine similarity of these two vectors is calculated using the cosine similarity formula. The specific calculation process involves first calculating the dot product of vectors A and B, i.e., a1×b1+a2×b2+…. Then, the moduli of vectors A and B are calculated separately. For example, the modulus of vector A is sqrt(a12+a22+…), and the modulus of vector B is sqrt(b12+b22+…). Finally, the dot product is divided by the product of the two moduli to obtain the cosine similarity. Furthermore, the cosine similarity is dynamically scaled based on the equipment operating environment parameters. Elevator operating environment parameters include frequency of use, building height, and floor range. For example, if an elevator is frequently used, the weights of certain wear-related features may be increased when calculating cosine similarity, thereby adjusting the cosine similarity. A tall building may also affect the weights of safety device-related features in the cosine similarity calculation, thereby achieving dynamic scaling based on the device's operating environment parameters.

[0108] Step S1434: Optimizing the feature encoder of the fault prediction neural network through an adversarial training mechanism so that the cosine similarity reaches a preset stability threshold.

[0109] Finally, in the elevator system's fault prediction neural network, the feature encoder is responsible for extracting features from various input data. During this process, an adversarial training mechanism is employed to optimize the feature encoder. This adversarial training mechanism involves two networks: a generator network (the feature encoder can be considered part of the generator network) and a discriminator network. The generator network aims to generate feature representations that achieve a cosine similarity close to a preset stability threshold, while the discriminator network attempts to distinguish whether the generated feature representations meet this requirement. For example, the initial cosine similarity calculated is 0.6, and the preset stability threshold is 0.8. Under the guidance of the generator network, the feature encoder adjusts the feature extraction method for standard diagnostic reports and emergency diagnostic reports, resulting in changes in the feature representations of the equipment health status baseline vector and the real-time operation and maintenance response feature vector. The discriminator network evaluates these changes and, if it deems that the requirements have not been met, provides feedback to the generator network for further adjustments. With continuous iterative training, the feature encoder is gradually optimized, and the cosine similarity continues to approach the preset stability threshold, thereby achieving the purpose of optimizing the fault prediction neural network and generating a feature alignment optimization target component. This component can reflect the degree of alignment between the standard diagnostic report and the emergency diagnostic report in the device feature latent space, which helps to improve the accuracy and reliability of the entire fault prediction system in diagnosing elevator system faults.

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

[0111] Step A110 : constructing a device operation state transfer learning framework, in which the device feature template under historical normal operating conditions is used as a reference anchor point.

[0112] In this embodiment, taking the ventilation system in a building as an example, a framework for equipment operating status transfer learning is constructed to better utilize historical data to optimize fault prediction. The ventilation system includes components such as fans, air ducts, and filters. By collecting long-term operating data from the ventilation system under historical normal operating conditions, an equipment feature template is constructed. This equipment feature template contains information on multiple aspects, such as the fan's speed range during normal operation, the stable air pressure range within the air duct, and the filter's resistance range at normal filtration efficiency. This data forms a comprehensive equipment feature template, which serves as a reference anchor point. This is like determining an origin in a coordinate system; subsequent calculations and analyses are based on this origin. For example, the fan's speed during normal operation is stable between 1000 and 1200 rpm. This speed range is part of the equipment feature template and serves as an important reference in the transfer learning framework.

[0113] Step A120 : calculating the deviation trajectory between the current prediction feature and the reference anchor point in the degradation trend space in each training iteration.

[0114] During the training of a ventilation system fault prediction neural network, each iteration generates current prediction features. These features reflect an estimate of the ventilation system's current operating state. For example, in one training iteration, the predicted fan speed is 1100 rpm, the duct pressure is 0.8 standard atmospheres, and the filter resistance is 10 Pascals. The reference anchor point has a fan speed range of 1000-1200 rpm, a standard duct pressure of 1 standard atmosphere, and a normal filter resistance range of 8-12 Pascals. The deviation trajectories of these predicted features from the reference anchor point are calculated in the degradation trend space. While the deviation in fan speed may be minor, the deviation in duct pressure indicates possible duct blockage or reduced fan efficiency. This deviation is not a simple numerical difference; rather, it represents a deviation in the degradation trend space that takes into account the overall operating logic of the ventilation system and the interrelationships between its components. For example, if the fan speed gradually decreases and the duct pressure continues to deviate from the standard value over multiple iterations, this deviation trajectory indicates a potential trend toward failure in the ventilation system. By calculating the deviation trajectories of the predicted features of multiple components from the reference anchor points, a comprehensive understanding of the changes in the operating status of the ventilation system can be obtained.

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

[0116] Model parameter regularization constraints are generated based on the dynamic rate of change of the ventilation system's deviation trajectory. For example, the fan speed deviation trajectory shows a decrease of 50 rpm every 10 iterations over a period of time, and the duct pressure deviation trajectory shows a decrease of 0.05 standard atmospheres every 10 iterations. Based on these dynamic rates of change, model parameter constraints for the fault prediction neural network can be determined. If the rate of change is too rapid, it indicates that the model may be overly sensitive to changes in the ventilation system's state or exhibit bias, requiring model parameter constraints. This constraint is incorporated into the composite training objective function, which originally optimizes the fault prediction neural network by comprehensively considering multiple factors (such as standard diagnostic reports, emergency diagnostic reports, and other relevant factors). Adding this model parameter regularization constraint based on the dynamic rate of change of the deviation trajectory allows the model to pay more attention to the actual operating state trends of the ventilation system during training, avoiding overfitting or incorrect predictions. For example, without this constraint, the model may focus too much on the prediction error at a certain moment and ignore the overall degradation trend of the ventilation system. After adding the constraint, the model can better adapt to the long-term operating status changes of the ventilation system while ensuring the overall prediction accuracy.

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

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

[0119] In ventilation system fault prediction, an online knowledge distillation mechanism is implemented to improve the performance and interpretability of the fault prediction neural network. Domain experts, drawing on years of experience and theoretical knowledge in ventilation system operation and maintenance, constructed a fault decision tree model as a teacher network. This fault decision tree model covers various possible ventilation system fault scenarios and their corresponding diagnostic paths. For example, the top-level node in the fault decision tree might be whether the ventilation volume is normal. If the ventilation volume is abnormal, the next node might be whether the fan is operating normally. If the fan is operating normally, the next node might be whether the air duct is blocked, etc. This hierarchical fault decision tree model can accurately determine the fault type and location based on the different symptoms of the ventilation system. Using it as a teacher network provides valuable prior knowledge for the fault prediction neural network.

[0120] Step B120: extracting the diagnostic path features of the teacher network in a typical fault scenario and generating a device status discrimination rule reinforcement signal.

[0121] When a typical failure 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 operating normally but the ventilation volume is zero, the teacher network will step by step determine the diagnostic path of the air duct blockage along the nodes of the fault decision tree. This diagnostic path contains multiple features, such as the fan speed is normal but the air pressure difference between the two ends of the air duct is too large. These diagnostic path features are extracted and converted into a signal to enhance the equipment status judgment rules. This signal contains the rule information on how to accurately judge the fault based on the status of each component of the ventilation system in the typical failure scenario of air duct blockage. For example, this signal may include specific judgment rules for determining air duct blockage based on parameters such as fan power, air pressure difference in the air duct, ventilation volume, etc. These rules can enhance the judgment ability of the fault prediction neural network in similar scenarios.

[0122] Step B130: injecting 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 explanatory diagnostic logic flow.

[0123] The attention fusion module plays a key role in the ventilation system's fault prediction neural network. It injects signals reinforcing the device state judgment rules into the decision layer of the fault prediction neural network. For example, when the fault prediction neural network determines whether a ventilation system has a duct blockage fault, the decision layer originally made its judgment based on the neural network's internal parameters and algorithms. Now, with the device state judgment rule reinforcement signal injected by the attention fusion module, the decision layer simultaneously considers the judgment rules extracted from the teacher network (fault decision tree model). For example, when determining duct blockage, the decision layer not only considers the relationship between duct pressure difference and ventilation volume learned by the neural network, but also considers judgment rules such as the reasonable range of duct pressure difference and ventilation volume under different fan power levels, obtained from the teacher network. This generates an interpretable diagnostic logic flow. In practical applications, if a duct blockage is diagnosed, not only can the conclusion be given, but the injected judgment rules can also be used to explain the specific values ​​and relationships between parameters such as fan power, duct pressure difference, and ventilation volume that led to the conclusion, thereby improving the credibility and understandability of the fault prediction.

[0124] The specific implementation steps of the online knowledge distillation mechanism include:

[0125] 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.

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

[0127] A knowledge transfer loss function is constructed based on the similarity, and the ability of the student network to capture the hidden fault characteristics of the equipment is dynamically adjusted based on the knowledge transfer loss function.

[0128] In this embodiment, a dual-channel comparative learning architecture is established during the equipment feature encoding stage of the ventilation system. 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 filtering principle of the filter. For example, according to the physical model of the fan, the air volume and air pressure of the fan have a specific relationship at different speeds. The teacher network channel uses this physical model knowledge to output a diagnostic decision path based on the physical model when receiving the operating data of the ventilation system (such as fan speed, air duct pressure, ventilation volume, etc.). For example, if the fan speed is normal but the ventilation volume is low, based on the physical models of the fan and the air duct, the teacher network channel may make a judgment 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) predicts a ventilation system fault, it generates predictions, such as whether the air duct is blocked or whether the fan is faulty. The feature similarity between the student network's prediction and the teacher network's decision path at the fault mode discrimination node is calculated. For example, for the fault mode discrimination node of duct blockage, the teacher network's decision path might determine the likelihood of duct blockage based on parameters such as duct pressure difference and ventilation volume. The student network also analyzes and predicts these parameters based on its learned model. The similarity between the two parameters is calculated. For example, if the teacher network determines an 80% probability of duct blockage based on a duct pressure difference greater than 0.5 standard atmospheric pressure and a ventilation volume less than 80% of normal levels, while the student network determines a 70% probability of duct blockage based on the same parameters, the feature similarity between the two at this fault mode discrimination node is calculated using a specific algorithm (such as cosine similarity or Euclidean distance).

[0130] A knowledge transfer loss function is constructed based on the calculated feature similarity between the student network's predictions and the teacher network's decision path at the fault mode discrimination node. This loss function reflects the degree of difference between the student network and the teacher network. A high feature similarity indicates that the student network's predictions are closer to the teacher network's decision path, resulting in a low loss function value; conversely, a high loss function value is observed. Based on this knowledge transfer loss function, the student network's ability to capture hidden fault characteristics of the equipment is dynamically adjusted. For example, a high knowledge transfer loss function value at the air duct blockage fault mode discrimination node indicates that the student network may not fully consider some hidden fault characteristics when determining air duct blockage (such as minor structural deformation within the duct that, while not completely blocking the duct, affects ventilation volume). By adjusting the student network's parameters, such as increasing the weight of features related to the duct's internal structure, the student network's ability to capture these hidden fault characteristics is improved, enabling the student network to more accurately identify ventilation system faults in subsequent predictions, particularly those that are difficult to detect.

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

[0132] Step C110 : 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.

[0133] In this embodiment, taking the elevator system in a building as an example, the device state evolution prediction module aims to predict the evolution of the elevator's health status in advance. When predicting faults in the elevator system, the fault prediction neural network generates intermediate feature outputs. These intermediate features contain operating status information for each elevator component. For example, for the elevator's traction motor, these intermediate features may include its vibration frequency and temperature change rate during operation; for the elevator car, these may include the car's sway amplitude and operating speed stability. A multidimensional health index curve is constructed based on these intermediate feature outputs. For example, the traction motor's vibration frequency is assumed to be 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 multidimensional health index curve. Simultaneously, the car's sway amplitude is also reflected in another dimension of the curve. If the car's sway amplitude gradually increases from the normal 0-1 mm to 2-3 mm, this will also be recorded on the multidimensional health index curve. In this way, the multidimensional health index curve integrates the status information of each key elevator component, forming an indicator that comprehensively reflects the elevator's health status.

[0134] Step C120 , synchronously optimizing the shared parameters of the equipment state evolution prediction module and the diagnosis module during the training process, so that the mutation points of the multidimensional health index curve form a spatiotemporal correlation with the abnormal patterns of the multidimensional fault feature distribution matrix.

[0135] During the training process of the elevator system, the equipment state evolution prediction module and the diagnosis module share parameters. For example, parameters for determining traction motor faults are used both in the diagnosis module to determine the presence of a fault and in the equipment state evolution prediction module to predict the evolution of the traction motor's health state. These shared parameters are optimized simultaneously to ensure a spatiotemporal correlation between the abrupt changes in the multidimensional health index curve and the abnormal patterns in the multidimensional fault signature distribution matrix. When the bearings inside the traction motor begin to wear, features associated with bearing wear (such as abnormal vibration frequency and elevated temperature) will show abnormal patterns in the multidimensional fault signature distribution matrix. Simultaneously, due to the effects of the shared parameters, a sudden change in the multidimensional health index curve will also appear at the corresponding time point. This change may be a sudden increase in vibration frequency or a rapid rise in temperature. By simultaneously optimizing the shared parameters, this spatiotemporal correlation accurately reflects the time of the elevator fault and the corresponding state change. For example, if at a certain moment, the car sway amplitude and the traction motor vibration frequency in the multidimensional health index curve show sudden changes at the same time, and the abnormal patterns of the car guide rails and traction motor are also shown in the multidimensional fault feature distribution matrix, it can be accurately determined that these two components may fail at the same time or have related faults around this time point.

[0136] In step C130 , the stage characteristics of the equipment degradation process are captured through a temporal convolutional network to generate timing trigger conditions for preventive maintenance decisions.

[0137] Elevator systems undergo different stages of degradation over long-term operation. A temporal convolutional network (TCN) analyzes elevator system operational data to capture the characteristics of these degradation stages. For example, in the early stages of degradation, some elevator components may only show minor wear and tear, which may be reflected in the operational data as small fluctuations in the traction motor's vibration frequency and occasional minor instabilities in the car's running speed. The TCN can identify these early degradation characteristics. As time passes, entering the mid-stage of degradation, more obvious characteristics may emerge, such as an increase in the frequency of traction motor temperature increases and increased car sway. The TCN performs convolution operations on operational data at different time points to capture the changing trends of these stage characteristics. Based on these captured stage characteristics, time-series trigger conditions for preventive maintenance decisions are generated. For example, if the temporal convolutional network detects that the fluctuation of the hoist vibration frequency exceeds the normal range more than a certain threshold number of times in a consecutive month, and the amplitude of the car shaking also tends to gradually increase, it can trigger preventive maintenance decisions, such as arranging inspections of the traction motor and calibration of the car guide rails, to avoid further degradation of the elevator and more serious failures.

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

[0139] Step D110: Configure a dynamic adversarial training strategy to generate an adversarial sample set that simulates abnormal device status.

[0140] In elevator systems, a dynamic adversarial training strategy is implemented to improve the robustness of fault prediction neural networks. First, an adversarial sample set is generated to simulate abnormal device conditions. For different elevator components, adversarial samples are generated by performing specific transformations on normal operating data. For example, for a traction motor, data such as vibration frequency and temperature during normal operation are known. Abnormal conditions are simulated by adding targeted noise or performing specific numerical transformations on this data. For example, periodic interference signals are added to the traction motor's vibration frequency data according to a specific algorithm to make the vibration frequency appear to be that of a fault (such as rotor imbalance). For car speed data, the slope of the data can be altered or random fluctuations can be added to simulate speed anomalies caused by guide rail wear or brake system failure. This generated adversarial sample set contains a variety of simulated elevator abnormality data and can be used for subsequent dynamic adversarial training.

[0141] Step D120 , alternately inputting real device data streams and adversarial samples in the adversarial sample set in each training iteration, and calculating the diagnostic robustness index of the fault prediction neural network in a composite disturbance environment.

[0142] During each training iteration of the elevator system fault prediction neural network, real device data streams and adversarial examples from the adversarial sample set are alternately fed into the training process. The real device data stream contains both normal and abnormal data from actual elevator operation. When fed with real device data streams, the fault prediction neural network performs fault prediction according to the normal prediction process. When fed with adversarial examples, since these simulate abnormal conditions, the neural network must accurately identify these anomalies. In this complex perturbation environment (both real data and adversarial examples with simulated anomalies), a diagnostic robustness metric is calculated. This metric reflects the stability and accuracy of the fault prediction neural network under various real and simulated conditions. For example, the diagnostic robustness metric can be measured by calculating the correct prediction ratio for real anomaly data and anomaly data from the adversarial sample set. If the neural network correctly predicted 90 out of 100 real anomaly data and 80 out of 100 anomaly data from the adversarial sample set, the diagnostic robustness metric can be calculated based on these two ratios.

[0143] Step D130 , adaptively adjusting the weight distribution ratio of each optimization component in the composite training objective function based on the diagnostic robustness index, and generating a neural network iterative mechanism with environmental adaptability.

[0144] For example, based on the calculated diagnostic robustness index, the weight distribution ratio of each optimization component in the composite training objective function is adaptively adjusted. The composite training objective function includes multiple optimization components, such as those related to standard diagnostic reports and those related to emergency diagnostic reports. If the diagnostic robustness index is low, it indicates that the fault prediction neural network, under its current training state, is insufficiently capable of handling abnormal situations. In this case, the weights of optimization components related to improving diagnostic accuracy may need to be increased. For example, if inaccurate predictions are found for abnormal data in adversarial examples, the weights of optimization components related to precise fault location information in standard diagnostic reports may need to be increased to allow the neural network to focus more on learning accurate fault location knowledge. This adaptive adjustment of weight distribution ratios creates an environmentally adaptable neural network iteration mechanism. This mechanism dynamically adjusts the training direction based on the performance of the fault prediction neural network at different training stages and when exposed to different data types (real data and adversarial examples), enabling the neural network to continuously adapt to the actual operating environment of the elevator system and improve its fault prediction capabilities in various situations.

[0145] Figure 2 FIG. 1 shows the hardware structure 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 provided by the embodiment of the present invention. Figure 2 As 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 data and / or instructions used by the intelligent system-based AI intelligent diagnostic system 100 to execute or use to complete the exemplary methods described in the present invention.

[0147] During the specific implementation process, one or more processors 110 execute 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 in the above method embodiment. 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 sending and receiving actions of the communication unit 140.

[0148] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned AI intelligent diagnosis system 100 based on the intelligent system. The implementation principles and technical effects are similar and will not be repeated 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-mentioned AI intelligent diagnosis method based on the intelligent system is implemented.

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

Claims

1. An AI intelligent diagnosis method based on an intelligent system, characterized in that: The method comprises: The operating data stream of the building's electromechanical equipment is collected through edge computing nodes, and the operating data stream is state-sequentially encoded to generate an equipment state sequence. The operating data stream includes vibration spectra, thermal imaging features, 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 potential equipment failure modes in preset diagnostic dimensions and abnormal confidence levels of associated components; Obtain standard diagnostic reports generated by the expert knowledge base and emergency diagnostic reports generated by on-site operation and maintenance records. The standard diagnostic reports contain precise fault location information and complex repair strategies, while the emergency diagnostic reports contain temporary disposal plans and unverified anomaly speculations. Constructing a composite training objective function based on the multidimensional 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 based on the target fault prediction neural network, fault prediction diagnosis is performed on the operating data stream of any input building mechanical and electrical equipment.

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 a reliability loss value for each verified diagnostic unit in the standard diagnostic report and an error correction value for each unverified diagnostic unit in the emergency diagnostic report, wherein the reliability loss value is a cross-validation result of the device state parameter within the diagnostic unit on the failure 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 device feature latent space, calculating the distribution difference between the two in the device health status discrimination space, and generating a feature alignment optimization target component; The diagnostic accuracy optimization objective component and the feature alignment optimization objective 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 the diagnosis unit, obtaining the credibility distribution of the fault mode corresponding to the device state determination node in the multi-dimensional fault feature distribution matrix; Based on the dynamic deviation between the equipment operating baseline parameters and the real-time monitoring parameters, a confidence decay function is constructed for each discriminant node in the diagnostic unit. 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, characterized in that: The method further comprises: Access maintenance decision records from the building equipment management system in real time to extract verified optimized diagnostic segments and temporary diagnostic segments awaiting review; The sample weights of the standard diagnostic reports and the emergency diagnostic reports are dynamically adjusted during the training process, and the weight coefficients of the optimized diagnostic segments increase with their effectiveness evaluation values ​​in actual operation and maintenance.

5. The AI ​​intelligent diagnosis method based on the intelligent system according to claim 2, characterized in that: The high-order representation 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 objective component include: Converting the standard diagnostic report into a device health status reference vector using 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 impact factor; Calculating 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 scaling the cosine similarity based on the device 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 the intelligent system according to claim 1, characterized in that: The method further comprises: Constructing a device operation state transfer learning framework, in which a device feature template under historical normal operating conditions is used as a reference anchor point; In each training iteration, calculating the deviation trajectory between the current predicted feature and 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 the intelligent system according to claim 1, 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 a device status discrimination rule reinforcement signal; The device state discrimination rule reinforcement signal is injected into the decision layer of the fault prediction neural network through the 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 during the device feature encoding phase, where 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 the hidden fault characteristics of the equipment is dynamically adjusted based on the knowledge transfer loss function.

8. The AI ​​intelligent diagnosis method based on the intelligent system according to claim 1, 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 points of the multidimensional health index curve and the abnormal patterns of the multidimensional fault feature distribution matrix form a spatiotemporal correlation; The stage characteristics of the equipment degradation process are captured through a temporal convolutional network, and the temporal trigger conditions for preventive maintenance decisions are generated.

9. The AI ​​intelligent diagnosis method based on the intelligent system according to claim 1, 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 according to any one of claims 1 to 9 above.

Images